A Social Theory of Knowledge

by

Boaz Miller

A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy Institute for the History and Philosophy of Science and Technology University of Toronto

© Copyright by Boaz Miller, 2011 A Social Theory of Knowledge

Doctor of Philosophy, 2011 Boaz Miller

Institute for the History and Philosophy of Science and Technology University of Toronto

Abstract

We rely on science and other organized forms of inquiry to answer cardinal questions on issues varying from global warming and public health to the political economy. In my thesis, which is in the intersection of philosophy of science, social epistemology, and science and technology studies, I develop a social theory of knowledge that can help us tell when our beliefs and theories on such matters amount to knowledge, as opposed to mere opinion, speculation, or educated guess. The first two chapters discuss relevant shortcomings of mainstream analytic epistemology and the sociology of knowledge, respectively. Mainstream epistemology regards individuals, rather than communities, as the bearers of knowledge or justified belief. In Chapter 1, I argue that typically, only an epistemic community can collectively possess sufficient justification required for knowledge. In Chapter 2, I present a case study in computer science that militates against the sociological understating of knowledge as mere interest‐based agreement. I argue that social interests alone cannot explain the unfolding of the events in this case.

Rather, we must assume that knowledge is irreducible to social dynamics and interests.

In Chapter 3, I begin my positive analysis of the social conditions for knowledge. I explore the question of when a consensus is knowledge based. I argue that a consensus is knowledge based when knowledge is the best explanation of the consensus. I identify

‐ ii ‐ three conditions – social diversity, apparent consilience of evidence, and meta‐agreement, for knowledge being the best explanation of a consensus. In Chapter 4,

I illustrate my argument by analyzing the recent controversy about the safety of the drug Bendectin. I argue that the consensus in this case was not knowledge based, and hence the deference to consensus to resolve this dispute was unjustified. In chapter 5, I develop a new theory of the logical relations between evidence and social values. I identify three roles social values play in evidential reasoning and justification: They influence the trust we extend to testimony, the threshold values we require for accepting evidence, and the process of combining different sorts of evidence.

‐ iii ‐ Acknowledgements

I owe many thanks to many people. I am fortunate to have had an active PhD committee, whose members’ different personas well complemented one another. I thank first and foremost my supervisor, Anjan Chakravartty, for his continuous help and support along the way from the very early stages of this project. I admire his excellent perceptive comments, and his insistence on conceptual clarity and transparent argument. Not only is Anjan a great philosopher who has given me great feedback, he has also been a highly efficient and dependable supervisor, who made the entire process of the dissertation run like an oiled clock. I thank Yossi Berkovitz, my perfectionist committee member, who does not know the meaning of the word ‘compromise’. Yossi knows only one way to do philosophy – the correct way, and he is stubborn. He has always pressed me to think about the big questions of philosophy, and how my work fits with respect to them. In our long, tedious and often loud arguments about my work and philosophy in general, he has taught me to aspire to meet the highest standards, even if, at the end of the way, I only partly agree with him about what they are. I thank my committee member Jim Brown for the helpful brain storming sessions I have had with him, and his accurate and precise comments on my work. Jim is the perfect person to bounce ideas with, and he has the unique ability to summarize in one sentence what my argument is about and what’s wrong with it. Above all, Jim is perhaps the nicest man I know, and he is always ready to help. Jim was the first person to tell me long ago that if was interested in the sort of questions I was, I might want to check out this field called social epistemology. Many people have commented on earlier parts of this dissertation. I thank Arnon Keren, Anat Leibler, Yanni Nevo, Yakir Levin, Shelly Kreiczer Levy, Omer Levy, Ori Aronson, Heather Douglas, Torsten Wilholt, Dave Matheson, Duncan Pritchard, Michael Lynch, Tamir Tassa, Marga Vicedo, Mark Solovey, Lucia Dacome, and Yves Gingras. I thank Liora Salter, who was in my specialist committee at the initial stages of my research project before it took a somewhat different direction, for her helpful suggestions and criticism. I thank my external reader Sandy Goldberg for his very

‐ iv ‐ thorough, detailed, and constructive report, and his helpful suggestions for improvement. I have been fortunate to be part of a vibrant student community in and outside IHPST. I especially thank my fellow young philosophers Isaac Record, Jacob Stegenga, and Eran Tal for long discussions over beer and wings, and very helpful feedback on earlier drafts. I also thank my friends at the institute and outside it – Vivien Hamilton, Delia Gavrus, Curtis Forbes, Mike Thicke, Michael Cournoyea, Becky Moore, Jai Virdi, Sebastián Gil‐Riaño, Ari Gross, Ellie Louson, Gwyndaf Garbutt, Chris Belanger, Ben Almassi, Victor Boantza, Allan Olley, John Turner, Sarah Scharf, Michelle Hoffman and everybody I might have missed who have endured with me in this PhD program. I thank the members of the Israeli community, International students, visiting scholars and other visitors in Toronto for their company and friendship, in particular Revital Goldhar, Josh Turner, Merom Kalie, Rayner Thwaites, Emily Hammond, Yaara Lamberger Kinar, Doron Kinar, Ayelet and Shai Lahat, Mohammed Wattad, Alon Harel, Nir Lefler, Oren and Libby Barak, Adi Folkman, Boaz Ben‐David, Efrat Shapir, Andrew Theobald, Israela Stein, Eitan Hess, Dror Harel, Nahshon Perez, Alma Gadot, Doron Pearl, Inmar and Gadi Givoni, Chen Lev Eshkolot, and everybody else I might have missed. Last, I thank my mother Leah, whom I suspect of secretly planning this PhD since I was a young boy, my father Uri for his support of every life choice I made, my brother Dror for deep philosophical discussions about science, and last, my wife, Meital Pinto, who has been my muse, inspiration, and role model in every possible way.

‐ v ‐ Table of Contents

INTRODUCTION...... 1 1. MOTIVATIONS AND AIMS ...... 2 2. METHODOLOGY ...... 5 3. SITUATING MY THEORY WITH RESPECT TO STS...... 8 4. OVERVIEW OF THE ARGUMENT...... 16 CHAPTER 1 EPISTEMIC COMMUNALISM DEFENDED...... 21 INTRODUCTION...... 21 1. THE THESIS OF EPISTEMIC COMMUNALISM AND HARDWIG’S ARGUMENT FOR IT ...... 22 2. THREE OBJECTIONS TO EPISTEMIC COMMUNALISM ...... 30 2.1. The Objection from Convergence of Multiple Confirmations...... 30 2.2. The Objection from Indirect Social Evidence ...... 32 2.3. The Objection from Reliability Indicators...... 34 2.4. Conclusion...... 37 3. THE ARGUMENT FROM EPISTEMIC IMPOSSIBILITY FOR EPISTEMIC COMMUNALISM ...... 37 4. RELIABILISM TO THE RESCUE OF INDIVIDUALISM? ...... 41 CONCLUSION ...... 47 CHAPTER 2 RELATIVISM AND THE LIMITS OF SOCIAL EXPLANATION: A PRIME EXAMPLE..... 49 INTRODUCTION...... 49 1. BETWEEN POPULARIZATION, SIMPLIFICATION AND DISTORTION ...... 54 2. ‘PRIMES IS IN P’ – NECESSARY SCIENTIFIC BACKGROUND ...... 62 2.1. Complexity Class P...... 62 2.2. Probabilistic and Deterministic Algorithms ...... 64 2.3. The Difference between PRIMES and IFP ...... 65 2.4. The RSA Encryption Algorithm...... 67 2.5. Conclusion...... 68 3. THE GENERAL PRESS COVERAGE OF ‘PRIMES IS IN P’...... 69 4. THE SHARED INTERESTS OF THE SCIENTISTS AND THE PRESS IN DISTORTION ...... 77 4.1. Visibility ...... 82 4.2. Recognition...... 83 4.3. Priority...... 85 5. POPULARIZATION AND DISTORTION REVISITED...... 86 CONCLUSION ...... 95 CHAPTER 3 KNOWLEDGE­BASED CONSENSUS...... 96 INTRODUCTION...... 96 1. KNOWLEDGE, LUCK AND INFERENCE TO THE BEST EXPLANATION...... 100 2. CONSENSUS, RATIONALITY AND IDEALIZATION...... 109 3. THE APPARENT CONSILIENCE OF EVIDENCE CONDITION ...... 111 3.1. First Approximation: Goldman’s Causal Independence Condition...... 111 3.2. Second Approximation: Tucker’s Epistemic Theory of Consensus ...... 114 3.3. A Challenge for Tucker: Solomon’s Accidental Aggregation Theory of Consensus...... 116 3.4. The Apparent Consilience of Evidence Condition ...... 123 4. THE SOCIAL DIVERSITY CONDITION ...... 127 5. THE META‐AGREEMENT CONDITION ...... 132 6. A COMPARISON WITH CRITICAL CONTEXTUAL EMPIRICISM...... 140 CONCLUSION ...... 145

‐ vi ‐ CHAPTER 4 WAS THE CONSENSUS ON BENDECTIN KNOWLEDGE BASED? ...... 147 INTRODUCTION...... 147 1. THE BENDECTIN TRIALS – GENERAL BACKGROUND ...... 148 2. WAS THE CONSENSUS ON BENDECTIN KNOWLEDGE BASED?...... 151 3. THE WIDER PICTURE – GENERAL LESSONS FROM THE BENDECTIN CASE STUDY ...... 161 CONCLUSION ...... 164 CHAPTER 5 THE ROLE OF VALUES IN EVIDENTIAL JUSTIFICATION...... 165 INTRODUCTION...... 165 1. THE UNDERDETERMINATION GAP‐FILLING MODEL AND ITS LIMITATIONS...... 166 2. SOCIAL VALUES AFFECT TRUST IN TESTIMONY...... 172 3. VALUES, EVIDENCE AND MOTIVATED REASONING...... 183 4. SOCIAL VALUES LOWER AND RAISE EVIDENTIAL THRESHOLDS...... 188 5. VALUES AFFECT THE RELATIVE WEIGHING OF DISCORDANT EVIDENCE...... 196 6. APPLYING THE FRAMEWORK TO EXISTING CASE STUDIES ...... 202 6.1. Values and Evidence in Dioxin Cancer Research ...... 202 6.2. Values and Evidence in the Gravity Waves Controversy...... 204 7. BROADER IMPLICATIONS FOR EPISTEMOLOGY AND PHILOSOPHY OF SCIENCE ...... 207 CONCLUSION ...... 210 CONCLUSION ...... 211 BIBLIOGRAPHY...... 217

‐ vii ‐

Introduction

Is human activity causing global warming? Is it safe for pregnant women to take antidepressants? Are antidepressants efficacious at all? Do gravity waves exist? Can they be detected? Is it safe to access bank accounts online? Are housing prices headed for a fall? What is the best food for newborn babies? Can dioxin cause cancer? Is

243112609 −1 a prime number? Can vaccinations cause autism? Is it safe to consume genetically engineered crops?

These are just some of the questions to which people in our world seek answers.

Sometimes they want answers just to satisfy their curiosity, but often they require them in order to make informed decisions, sometimes of vital importance, about which course action to take. In the vast majority of cases, they cannot find answers to these questions on their own. Rather, they must turn to experts, science, and other specialized fields of inquiry.

When they do so, a new family of meta‐questions arises: Who are the experts in a given domain? Which experts are reliable? Does the expert advice represent a mere opinion, an educated guess, or genuine knowledge? Who should be trusted when experts disagree? When a scientific community is split, is the majority more likely to be correct? Does the fact that all relevant specialists seem to agree indicate that they are more likely to be right? How do biases, interests and social values affect specialists’ opinions and to what extent?

While the first family of questions belongs to science and other specialized fields, the second family of meta‐questions belongs to the realm of philosophy of science and social epistemology – the branch of epistemology that studies the social dimensions

‐ 1 ‐

of knowledge. When we turn to the philosophy of science and social epistemology, however, we often find them struggling to provide us with systematic and satisfactory answers to such questions. While these fields study science, expertise, and society, their discussion of these issues seldom directly answers questions about the trustworthiness of the sources of knowledge to which we turn. In this dissertation, I make a step toward remedying this situation. I lay a foundation for a social theory of knowledge that is specifically attuned to these sorts of questions, and can rise to the challenge they pose.

This dissertation lies in the intersection of philosophy of science, social epistemology and science and technology studies (STS). In section 1 of this introduction,

I review my motivations and aims and discuss them in the context of the aims of traditional epistemology. In section 2, I review my methodology and situate my dissertation with respect to similar projects in the philosophy of science. In section 3, I discuss the empirical data on which I rely and situate my dissertation project with respect to the discourse of STS. In section 4, I provide a brief overview of the chapters and the overall argument.

1. Motivations and Aims

Epistemology has traditionally undertaken three projects. The first is the analysis of knowledge. This project has flourished in the twentieth century, especially after the publication of Gettier’s (1963) paper, which exposed difficulties with the traditional analysis of knowledge as justified true belief.1 The second project is battling scepticism.2

This project has occupied a central place in mainstream epistemology, and the battle has proven to be fierce and bloody. The third project is developing criteria for

1 See Steup (2008) for an overview. 2 See Klein (2009) for an overview.

‐ 2 ‐

evaluating beliefs that can guide people in acquiring true beliefs and avoiding false beliefs. While this project troubled enlightenment philosophers, such as Locke,

Descartes and Hume, contemporary epistemology has somewhat stepped away from it, focusing more on the first two projects. A fourth relatively new project, which has been engaged mostly by social epistemologists in the last few decades, is studying how knowledge is connected to social structures and power relations.3

These four projects are interconnected. An adequate analysis of knowledge should rule out scepticism, when scepticism is inappropriate, and give us a good idea of how to evaluate the beliefs we hold. The relations between these projects, however, are not straightforward. In particular, the analysis of knowledge is not directly translatable into evaluative criteria for our beliefs. While contemporary epistemologists have heavily engaged with the analysis of knowledge and justification, they have much less engaged with the project of the epistemic evaluation of our beliefs. As Schickore (2009,

94) notes:

…analytic epistemologists seek to clarify what it means to say that S’s belief that p is justified, but they do not formulate criteria for the evaluation of whether the belief that p is, in fact, justified. According to reliabilist interpretations of justification, for example, the belief that p is justified if and only if it is produced by a reliable procedure. According to deontologist interpretations, the belief that p is justified if and only if the believer has not violated any of her epistemic duties. And evidentialist interpretations suggest that the belief that p is justified if and only if it fits the knower’s evidence. But reliabilists do not analyse concrete instances of reliable processes in science; deontologists do not spell out scientists’ epistemic duties in a given case; and evidentialists leave open what appropriate ‘justifiers’ are and how claims to knowledge should fit them.

In recent years, then, epistemology has advanced our understanding of the notions of knowledge and justification, but many questions that belong to the realm of the third project of evaluating our beliefs have not been satisfactorily answered.

3 See Goldman (2002a), Longino (2008) and Grasswick (2008) for an overview.

‐ 3 ‐

Because we acquire most of our beliefs from the testimony of others and from social institutions such as science that are responsible for the production of knowledge, social epistemology – the branch of epistemology that engages with the fourth project of studying the social dimensions of knowledge – provides us with fruitful resources for developing criteria to evaluate our beliefs. The main task of my dissertation, then, is to make a contribution toward developing a social theory of knowledge that can help us evaluate our beliefs. Such a theory will provide criteria we can use to evaluate our beliefs in some cases, and point out social factors and considerations that are relevant to formulating such criteria in other cases.

My dissertation tries to find out in what social circumstances our beliefs or theories amount to knowledge. It regards ‘knowledge’ as a success term that corresponds, as I will explain in Chapter 4, to a privileged form of belief or acceptance.

That is, it tries to identify social epistemic indicators for the existence of knowledge, as opposed to mere opinion, speculation or unwarranted beliefs in concrete case studies.

My project is similar in aims to other theoretical frameworks in social epistemology such as Social Empiricism (Solomon 2001) and Critical Contextual Empiricism (Longino

2002), which address the question of under what social conditions a theory adopted by an epistemic community is epistemically justified. While my framework shares some of the aims with these frameworks, it makes different assumptions and may differ in the conclusions it reaches in particular cases.

Because the four projects of epistemology are interconnected, the contribution of this dissertation will not be confined to the third and fourth projects. Rather, it will also contribute to the analysis of knowledge and the debate about scepticism. In fact, as

‐ 4 ‐

I will argue, the first steps in devising a framework that can help us evaluate our beliefs are revising the existing individualistic conceptions of knowledge in analytic epistemology, and rejecting a certain form of scepticism that is commonly associated with social constructivism.

The main aim of this dissertation project, then, is it to make a major contribution toward developing a social theory of knowledge that can guide us in acquiring true and warranted beliefs. In the next section I will discuss the methodology I employ in order to achieve this aim.

2. Methodology

In light of the aims I have set out for this dissertation, the methodology I will employ is normative naturalism. Normative naturalism, as its name suggests, stands on two pillars. The first pillar is naturalized epistemology. Naturalized epistemology is a family of views according to which the theory of knowledge is continuous with other theories about the constitution of the natural world, and the claims it makes about knowledge are to be adjudicated in the same manner claims in other naturalized research programmes, such as science, are adjudicated. This inter alia means that empirical evidence should play a role in adjudicating between claims about knowledge (Laudan

1996, 155). Naturalized epistemology need not exclude methods such as conceptual analysis or thought experiments, because they are legitimate methods in the sciences as well, but it does insist that empirical evidence be used as well.

The second pillar of normative naturalism is normativity. An obvious difficulty arises in naturalized epistemology with respect to the aspiration of traditional epistemology to make normative claims. Some proponents of naturalized epistemology

‐ 5 ‐

want to dispense with normativity altogether, as they believe normative recommendations have no place in an empirical research programme. Normative naturalism holds a different view, which will be the one I adopt here. Normative naturalism regards norms as empirical conjectures or hypothetical imperatives, linking means and ends. As such, empirical data about the ways in which knowledge is produced, how beliefs are acquired, and how certain methods have been successful or unsuccessful in achieving epistemic aims in the past are all relevant for coming up with normative recommendations (Laudan 1996, 156‐7).4

We can distinguish between three types of normative epistemological accounts, which I call directive, prescriptive, and evaluative. Directive normative accounts concern the question of what the aims of science or inquiry in general ought to be. Among the proposed aims we find truth simpliciter (Goldman 1999), significant truth (Kitcher

2001), empirical adequacy (van Fraassen 1980), explanatory coherence (Thagard

1992), empirical success (Solomon 2001), and practical intervention in pressing social problems (Harding 2002). Other accounts are pluralistic about the aims of science and allow them to change in time (Laudan 1984; Longino 2002).

Prescriptive normative frameworks concern the question of how the aims of science and inquiry in general ought to be achieved, thus they make recommendations about the proper conduct of inquiry. Much of the discussion of scientific method in the philosophy of science belongs to this genre. Among the proposed methods we find the

4 Naturalized epistemology comes in many flavours. Other than the claims I have made in these paragraphs, I am not committed to other views that are commonly associated with it. In particular, I am not committed to Quine’s (1969) recommendation of replacing traditional epistemology with cognitive psychology. Similarly, while I am sympathetic to Laudan’s view of normative naturalism as a philosophical methodology, I am not committed to his particular strand of normative naturalized epistemology, which makes further claims about the aims of science and the appropriate ways to achieve them.

‐ 6 ‐

method of conjectures and refutations (Popper 1968), the methodology of scientific research programmes (Lakatos 1970) and Bayesianism (Howson & Urbach 2006). What is common to these frameworks is that they all offer prescriptions about the ways in which scientists ought to reason and conduct inquiry. Prescriptive social epistemologies include Critical Contextual Empiricism (Longino 1992; 2002), which specify norms of inquiry required to achieve objectivity, and Fricker’s theory of epistemic injustice

(2007), which specifies the kind of intellectual and moral virtues a person ought to cultivate in order to acquire warranted beliefs, both in science and in general.

Last, normative evaluative accounts, such as my own, concern questions of how good a given epistemic state of affair is, for example, how good the beliefs we hold or the theories we accept are, or how efficient the distribution of research efforts is in an epistemic community. Rather than making recommendations for the conduct of inquiry, namely telling scientists how to act, as prescriptive accounts do, evaluative accounts devise standards to assess the product of inquiry. Social Empiricism (Solomon 2001) is an example of such a normative evaluative social epistemology. It seeks to answer questions such as whether the distribution of research efforts in a community is rational.

These three philosophical projects focus on different questions, but they are interrelated. Directive accounts need to show that the aims of inquiry they specify are achievable, prescriptive accounts should posit some aims that their recommended conduct of inquiry is designed to achieve, and both of them need to explain how to judge if the aims are satisfied. Evaluative accounts also need to posit some aims, and of course, one way by which they can evaluate beliefs and theories is to examine whether they

‐ 7 ‐

were produced by following certain recommended procedures. Therefore, although my primary motivation is evaluative, I will touch on questions of the other projects as well.

My framework regards knowledge, understood a privileged form of belief or acceptance, as an aim of science and inquiry in general. This does not mean that knowledge is the only or main aim of science. I neither deny that science may have merely instrumental or practical aims, such as enhancing our ability to manipulate and control nature, developing better technology, or alleviating famine and disease,5 nor do

I argue that such aims are less important than knowledge. Rather, I note that in many cases we do want science and inquiry to result in knowledge, and in those cases my theory can help us evaluate whether our theories and beliefs do in fact amount to knowledge.

3. Situating My Theory with Respect to STS

Normative naturalism stresses the importance of empirical evidence in substantiating epistemic claims. I will use three types of evidence – primary sources, historical case studies, and contemporary case studies. Most of the examples will be contemporary case studies from the field of STS, which brings me to discuss how my theory fits within this field. I will start with a brief history of the field, distinguish between two types of investigations that are conducted in it – theorizing and empirical case studies, and situate social epistemology and my own account in particular with respect to two other prominent theoretical approaches to knowledge in STS – social constructivism and heterogeneous constructivism.

5 For example, scientists may develop an effective treatment for curing a certain disease, and other than knowing that it is effective, they may not know or aim at knowing why it is effective, what the relevant causal mechanism is, which of the components of the treatment are necessary and which ones are redundant, etc.

‐ 8 ‐

STS – an acronym for ‘science and technology studies’ or ‘science, technology and society’ – is an interdisciplinary field that studies the social processes and outcomes of science and technology. Historically, STS developed from a family of views commonly called the Strong Programme in the Sociology of Knowledge. The Strong Programme was a new paradigm in the sociology of scientific knowledge that emerged in the 1970s.

Before the Strong Programme, the sociology of scientific knowledge addressed issues such as the social norms under which science functions optimally (Merton 1973), or tried to come up with unified organizing principles, such as scientists’ desire for recognition, to explain what drives scientific research (Hagstrom 1965). Proponents of the Strong Programme resented the fact that such work confined itself to the conduct of research and refrained from giving social explanation to the content of scientists’ beliefs. Similarly, they resented traditional history of science, in which they identified

Whiggish biases, namely a tendency to explain how the progression of scientific knowledge to its present state was inevitable. They also disapproved of what they saw as its tendency to invoke social factors only to explain erroneous beliefs, assuming that true and rational beliefs are self‐explanatory. Inspired inter alia by Kuhn’s The Structure of Scientific Revolutions (1962/1970), which they took to say that theory change cannot be explained rationally, they proposed a sociological framework that explains seemingly true, false, rational and irrational beliefs citing the same causes, including social causes, and exposes the contingent points in the formation of knowledge in which it could have taken different routes.6

6 There is an abundance of literature in and about the Strong Programme. For a theoretical manifesto of the Strong Programme see Bloor (1991), and Barnes & Bloor (1982). For a critique of the traditional history of science and the sociological alternative to it, see Collins (1982). For influential historical work in the spirit of the Strong Programme, see Shapin & Schaffer (1985), and Shapin

‐ 9 ‐

While STS has emerged from the Strong Programme in SSK, and still exhibits strong alliance to it, it would be wrong to equate the two. From its inception, STS has been wider both in its interests and its theoretical approaches. It has sought not only to explain how social factors influence scientists’ beliefs and theory acceptance, but all aspects of science and technology as social endeavours, their place in society, and their relations to other social systems such as law and the state. As Sheila Jasanoff, one of the founders of the field, puts this:

…the emerging field of science and technology studies (S&TS) has adopted as its foundational concern the investigation of knowledge societies in all their complexity: their structures and practices, their ideas and material products, and their trajectories of change. Growing from many disciplinary roots— including history, philosophy, sociology, politics, law, economics, and anthropology—S&TS today encompasses a rich tapestry of theoretical and methodological perspectives, all specifically directed toward investigating the place of science and technology in society (Jasanoff 2004, 2).

A look at a contemporary introduction to STS (Sismondo 2010) and the most recent authoritative Handbook of STS (Hackett et al. 2007), reveals that STS today is pluralistic, even eclectic, with respect to the theories on which it draws. In addition to SSK, among the theoretical frameworks in STS we find communication theory, political theory, feminist theory, anthropological theory, critical theory, and philosophy of science.

We can distinguish between two types of investigations that are carried out in

STS. The first is empirical research, which typically takes the form of detailed analyses of case studies. Such analyses typically draw on one or more theoretical frameworks as their analytical toolbox. Their primary interest is understanding the case study at hand.

They are not necessarily committed to the theories on which they rely in their entirety.

(1994). For a philosophical critique of the Strong Programme, see Laudan (1996) at 183‐209, and Brown (1989; 1994; 2001). For a history and a critique of STS from the perspective of contemporary history of science, see Daston (2009).

‐ 10 ‐

Rather, they selectively use analytical tools from various theories which they find useful. The second type of investigation, to which my account belongs, is theorizing.

Here the main aim is to come up with consistent theories of knowledge, and the role of the case studies is to lend support to the theory. While the distinction is not always clear‐cut, it is useful to note these two separate types of inquiry and the different roles empirical data and theory play in each type.

How are STS and social epistemology related? Social epistemologist Alvin

Goldman wants to subsume STS under the heading of social epistemology. He regards

STS as a branch of social epistemology that deals with science. He admits, though, that most STS researchers would not self‐identify as social epistemologists (Goldman 2002a,

197). In my view, it would be more correct to regard social epistemology as one of the theoretical frameworks used in STS. This view accords with social epistemologists’ conception of social epistemology as a theoretical framework that gives a better account than SSK of much of the empirical work conduced in STS (Kitcher 1994; Goldman 1994;

Fricker 1998; Lipton 1998, 4‐14). Social epistemology of science is said to be the philosophical and normative wing of STS, and a bridge between STS and philosophy of science and general epistemology, which sometime tends to indulge in armchair abstractions (Longino 2002; Solomon 2007a). I share this view and regard my account as part of this project.7

As we have seen, the most dominant theoretical framework in STS is the Strong

Programme in SSK. It is associated with constructivism and epistemic relativism, and almost all work in STS shows some allegiance to these notions. This is due partly to

7 An exception to this rule is Kusch’s communitarian epistemology (2002) which is a descriptive project in the spirit of the Strong Programme that aims at giving a conceptual analysis of notions such as knowledge, objectivity, and truth in terms of social agreement between agents.

‐ 11 ‐

historical reasons, as STS has emerged from SSK, but also, and perhaps mainly, because of its subject matter. STS frequently studies science in the making, in the stage before claims can be decisively labelled as true or false. It also studies controversies, in which the actors debate which claims are true and false. It wants to explain why certain claims gain trust, and why certain beliefs are held by some people despite being regarded by other people as false. The traditional analytic categories of truth and falsity are mostly unhelpful in dealing with these issues. A notion of knowledge as constructed and institutionally maintained, and an approach that does not prejudge certain claims as true and false, seems more appropriate for analyzing such case studies (Jasanoff 1996,

267).

A distinction exists between two types of constructivism: social and heterogeneous (Hess 1997, 82‐3; Sismondo 1996, 70‐1). Social constructivism is associated with three commitments. The first is contingency – the view that current accepted facts depend on certain social historical circumstances, and that they could have been different had the circumstances been different. To say that X is socially constructed is to say that ‘X need not have existed, or need not be at all as it is. X, or X as it is at present, is not determined by the nature of things; it is not inevitable’ (Hacking

1999, 6). The second commitment is nominalism – the view that the natural world has no unique internal structure, and hence no one preferred representation. The third commitment is external explanations of stability – the view that the stability of our theories and beliefs is explained mostly by the social interests they serve and the social institutions that maintain them, rather than by epistemic factors such as their truth or rationality (Hacking 1999, 68‐99).

‐ 12 ‐

My major engagement with social constructivism will be in Chapter 2, where I present a case study that militates against the third commitment, and argue that epistemic factors do in some cases explain stability and change in our beliefs. Hence the appropriate task of the social epistemologist is to draw a principled distinction between cases in which they do and do not.

Heterogeneous constructivism differs from social constructivism in that it denies any significant distinction between human and non‐human entities. The most developed theory in this genre is Actor Network Theory (ANT), which is very commonly drawn on in STS. Let me now briefly explain why despite its popularity, I will not substantively engage with it in my dissertation.

ANT emerged from social constructivism. In their earlier work, Latour and his colleagues worked within the framework of social constructivism and expressed similar views to that of the Strong Programme. In their later work, they changed their views and adopted the supersymmetry thesis, according to which accounts of science should be symmetrical with respect to human and non‐human entities, as there is no important distinction between them (Callon & Latour 1992). Symbolic of this theoretical shift is the fact that when the second edition of Latour and Woolgar’s 1979 book Laboratory

Life : The Social Construction of Scientific Facts came out in 1986, the word ‘social’ was omitted from its subtitle.

In addition to supersymmetry, ANT is committed to controversial metaphysical claims. It states that properties of objects exist only within a particular network with other objects. The nodes of such networks are human and non‐human actors. For

Latour, the (only) task of a scholar who is interested in studying scientific theories or

‐ 13 ‐

artefacts is to track the formation of the network of which they are part. That is to say,

ANT regards the world as composed of objects with no intrinsic properties whatsoever – only relational properties, which emerge in a specific network. For example, according to a traditional historical narrative, when in 1882, Koch discovered the bacteria that cause tuberculosis, he discovered bacteria that had already existed, and had always been responsible for the disease. Latour denies this. He argues that these bacteria were constructed in 1882 within a network of which Koch was part. They did not independently exist before that time. Consequently, King Ramses II, who lived about

3,000 years before Koch, could not have died from tuberculosis, as scientists now argue, because he was part of a different network of which tuberculosis and the entity that has the property of causing it were not part (Latour 2000).

Not surprisingly, while popular in some circles, Latour and ANT have also been heavily criticized by philosophers and sociologists alike. As for supersymmetry, it has been argued that the fact that humans are language users and rule followers is relevant, significant, and justifies treating them differently from non‐humans (Bloor 1999; Collins

2010). Furthermore, attempting to treat humans and non‐humans alike, ANT practitioners describe non‐humans as interacting with humans socially. This amounts to an indefensible anthropomorphizing of nature (Brown 2001, 134‐5). It is also argued that ANT metaphysical claims do not hold up to serious philosophical scrutiny (Brown

1994, 41‐59), and that the theory lacks consistency and conceptual clarity (Gingras

1995). ANT has also been criticized for failing on its own terms. It is not able to account for the fact that when instruments are taken out of the network of instrument builders and placed in a new network of instrument users, they preserve their original function

‐ 14 ‐

(Tal 2008). I find these criticisms mostly persuasive, and ANT philosophical assumptions remote from the ones I take as a starting point for the development of my own work here.

Leaving the philosophical problems aside, in STS practice, when ANT is used to analyze concrete case studies, it often collapses into a traditional and philosophically non‐illuminating historical narrative. When ANT accounts are purged of the metaphorical anthropomorphizing language they use to describe inanimate objects – for example of objects ‘forming alliances’ with people and then ‘betraying’ them (Latour

1987, 123), we are left with traditional and conservative historical narratives about objects not behaving as previously expected. In many other cases, scholars pay a lip service to ANT, but use a thin version of ANT that turns it into Strong‐Programme style social constructivism, where humans are the main actors and non‐humans play no significant part.8 As it is practiced, then, rather than as it is preached, my discussion of the shortcomings of social constructivism is applicable to ANT as well.

The role of STS case studies in my dissertation, then, is twofold. First, they provide rich and invaluable empirical data to be used as evidence to substantiate my theoretical claims. Second, they will be used to expose explanatory lacunas and shortcomings in existing STS theories, and to support my own proposed theory of

8 Here is a typical example. Alatout (2009) draws on ANT to track changing theories about the available quantity of drinking water in Palestine before and after 1948. Before 1948, when Palestine was under a British mandate, the British authorities restricted Jewish immigration to it based on the land’s capacity to sustain its inhabitants. Therefore, so Alatout argues, Zionist water scientists, who supported Jewish immigration, made optimistic estimations of the available amount of drinking water. After the state of Israel was formed in 1948 and experienced mass immigration waves, concerns were raised about its capacity to take in masses of people, and scientists’ estimates became much more conservative. Despite Alatout’s lip service to ANT, this is entirely a story about human actors and their political interests. As an actor, water itself has no active part in the network he depicts.

‐ 15 ‐

knowledge. Having laid out my theoretical approach and expected contribution, in the next section I will provide an overview of the argument chapter by chapter.

4. Overview of the Argument

The first two chapters discuss the shortcomings of analytic epistemology and SSK, respectively. Mainstream analytic epistemology regards individuals, rather than communities, as the bearers of knowledge or justified beliefs. Challenging this view,

Hardwig argues that laypersons and experts alike do not possess justification in the form of personally available direct evidence for most of their knowledge. A possible consequence of this claim is that only an epistemic community, whose members must trust each other, collectively possesses such justification. This means that the knowing subject is a community, rather than an individual, as traditionally thought in epistemology, and that individual knowledge depends on irreducible trust.

In chapter 1, I defend this view and provide a novel defence of epistemic communalism. Some social epistemologists have tried to resist epistemic communalism by arguing that individuals qua individuals possess sufficient indirect epistemic justification for their beliefs to count as knowledge. They try to show in different ways that while individuals cannot possess direct evidence for most of their beliefs, they can possess indirect evidence. I argue that these objections fail because the type of indirect evidence they identify is not good enough for knowledge. I further argue that even if we grant that such evidence is good enough for knowledge, individuals qua individuals cannot usually possess it in sufficient quantity for knowledge – only an epistemic community can collectively do that. Put differently, I argue that indirect justification is as dependent on irreducible trust as direct justification. Hence, only a community can

‐ 16 ‐

jointly possess it. I conclude by arguing that my argument applies to evidentialist and reliabilist theories of knowledge and justification alike.

If individualistic analytic epistemology is not apt for the challenge of addressing the social dimensions of knowledge, can SSK provide the answer? Crudely speaking, SSK constructivist epistemology analyzes knowledge and explains its stability largely in

terms of social agreement that serves various actors’ interests. In Chapter 2, I discuss a case study in computer science that suggests that SSK epistemology is unsatisfactory. I analyze the general media’s distorted coverage of the discovery of the AKS algorithm for identifying prime numbers in computer science in 2002. I argue that social interests alone cannot explain the unfolding of the events following this discovery. I argue that we must assume that knowledge is not reducible to social dynamics and interests if we want to explain why, although the scientists in this case published their discovery in the popular press, which distorted its meaning and implications, their claims gained the status of knowledge among their peers. I argue that this case study exposes a general weakness in existing epistemological assumptions underpinning current sociological theories of knowledge. Thus, it seems that neither analytic epistemology nor SSK gives a satisfactory social theory of knowledge.

After discussing the shortcomings of analytic epistemology and SSK epistemology, I begin my positive analysis of the social conditions for knowledge. I explore the question of the relations between knowledge and consensus. Scientific consensus is widely referred to in public debates as a social indicator of the existence of knowledge. For example, the National Institute of Health (NIH) and the

Intergovernmental Panel on Climate Change (IPCC) are in the business of formulating

‐ 17 ‐

expert consensus statements, which purport to provide authoritative answers to a variety of disputed questions. Additionally, the existence of consensus in a scientific community is used as a resource for arbitrating between rival expert testimonies in legal trials and making public policy decisions. However, it is far from clear that such deference to consensus is always justified. The existence of agreement is a contingent fact, and scientists may reach a consensus for all kinds of reasons, such as fighting a common foe or sharing a common bias. Scientific consensus, by itself, does not necessarily indicate the existence of knowledge.

I address the question of when a consensus is knowledge based. I argue that a consensus is knowledge based when knowledge is the best explanation of the consensus. I identify three conditions – social diversity, apparent consilience of evidence, and meta‐agreement, for knowledge being the best explanation of a consensus. My theory of knowledge‐based consensus resolves tensions between

Solomon’s Social Empiricism (2001) and Tucker’s (2003) theory of the epistemic significance of consensus. I argue that it overcomes the difficulties with the Critical

Contextual Empiricism (Longino 2002) analysis of consensus.

In Chapter 4, I illustrate my argument by analyzing the controversy in the 1980s and 1990s over whether the drug Bendectin caused birth defects. In the 1980s and early 1990s, there was a series of mass tort trials in U.S. Federal Courts involving birth defects allegedly caused by a drug called Bendectin, which was manufactured by Merell.

In the early 1980s courts tented to rule for the plaintiffs, but from the mid 1980s they started ruling against the plaintiffs on account of not showing causation. In the Daubert

(1993) decision, the U.S. Supreme Court ruled that the judge should act as a gatekeeper

‐ 18 ‐

to exclude unreliable or irrelevant evidence. Daubert legitimated district and circuit courts’ practice of excluding in vivo and in vitro studies as irrelevant or unreliable, ruling that epidemiological studies are necessary for showing causation. On a traditional narrative, the change in courts’ views reflected an emerging scientific consensus. Courts were merely passive assimilators of scientific knowledge . My analysis reveals that this consensus was not knowledge‐based, because it was not the product of a consilience of evidence, but rather it was achieved by the de facto termination of further in vivo and in vitro studies, due to the courts’ exclus ion of them as reliable scientific evidence in the mid 1980s. This occurred even though at that time, their results were still inconclusive. I argue that because the consensus was not knowledge‐ based, courts could not legitimately justify their decisions by deferring to it.

In chapter 5, I continue my analysis of the social conditions for knowledge. I offer a new theory of the logical relations between evidence and social values, such as political views and ideologies. The value‐ladeness of science is relatively uncontroversial among historians and philosophers of science. It is commonly argued that social values “fill the gap” of underdetermination of theory by evidence, namely social values direct our choice between two or more empirically adequate theories. I identify additional roles social values play in evidential reasoning and justification.

Drawing inter alia on research in experimental psychology, I argue that social values not only fill the gap between theory and evidence, but they also influence the trust we extend to testimony, the threshold values we require for accepting evidence, and the process of combining different sorts of evidence. My theory can explain, from an epistemic perspective, rather than a psychological perspective why a community of

‐ 19 ‐

smokers, for example, is able to find the same evidence about the dangers of smoking less persuasive than a community of non‐smokers. The upshot of this theory is that when we want to know whether our theories and beliefs are sufficiently supported by evidence, we must consider the possible influence of values on our judgment in the ways I have identified.

‐ 20 ‐

Chapter 1

Epistemic Communalism Defended

Introduction

The analysis of knowledge has been a central project of twentieth century epistemology.

With a few notable exceptions, and unlike other fields that are also devoted to the study of knowledge, such as feminist epistemology, the sociology of scientific knowledge and philosophy of science, mainstream analytic epistemology regards individuals, rather than communities, as the bearers of knowledge or justified beliefs.

In this chapter I argue that individualistic conceptions of knowledge cannot adequately characterize most of a normal individual adult’s beliefs that are commonly regarded as knowledge. I develop and defend John Hardwig’s argument for epistemic communalism from the irreducible role of trust in knowledge. I argue that the justification for most of a person’s beliefs is jointly possessed by the members of the epistemic community to which she belongs. I propose a communalist‐evidentialist conception of knowledge, and a reliabilist counterpart to it. I argue that only such communalist conceptions, which take into account the social dimensions of epistemic justification, can successfully engage with pressing real‐world sceptical worries that arise with respect to our beliefs.

This chapter contains five sections. In section 1, I present Hardwig’s argument for epistemic communalism and articulate the evidentialist conception of epistemic communalism it tacitly presupposes. In section 2, I refute three objections to communalism that try to show that even if individuals cannot have justification for their beliefs in the form of personally available direct evidence, they can still have indirect

‐ 21 ‐

evidence. In section 3, I argue that even if we grant these objections, it is impossible for individuals qua individuals to possess sufficient indirect evidence for their beliefs. In section 4, I argue that although the debate about communalism has been framed in evidentialist terms, individualism is not a viable option for reliabilists either for similar considerations. Reliabilist must acknowledge that the reliability of most of a subject’s beliefs extends far beyond that subject into the epistemic community of which she is a member.

1. The Thesis of Epistemic Communalism and Hardwig’s Argument for It

While epistemic communalism has not become part of mainstream analytic epistemology, there are several interestingly different communalist epistemologies

(Adler 2005; Kusch 2003; Longino 2002; Welbourne 2001; Nelson 1993). They all try to show in different ways that a communal account of knowledge better characterizes sociological or folk understandings of knowledge, which talk, for example, about how knowledge is transferred and shared. These accounts typically break away from the traditional analysis of knowledge in other ways as well. For instance, they do not regard knowledge as a species of belief. Traditional analytic epistemologists may therefore find such accounts objectionable for various reasons. Hardwig’s argument for epistemic communalism, on the other hand, starts with the traditional analysis of knowledge and shows that it necessitates that the knower be a community. While I am sympathetic to these other communal accounts of knowledge, I will focus on Hardwig’s argument because it has most in common with mainstream epistemology.

Hardwig begins his argument with the following observation:

‐ 22 ‐

I find myself believing all sorts of things for which I do not possess evidence: that smoking cigarettes causes lung cancer, that my car keeps stalling because the carburetor needs to be rebuilt, that mass media threaten democracy, that slums cause emotional disorder, that my irregular heart beat is premature ventricular contraction, that students' grades are not correlated with success in the nonacademic world, that nuclear power plants are not safe (enough)... That list of things I believe, though I have no evidence for the truth of them, is, if not infinite, virtually endless. And I am finite. Thought I can readily imagine what I would have to do to obtain the evidence that would support any one of my beliefs, I cannot imagine being able to do so for all my beliefs. I believe too much; there is too much relevant evidence (much of it available only after extensive, specialized training); intellect is too small and life is too short. (Hardwig 1985, 335).

Hardwig notes that in cases in which a layperson relies on expert testimony, he often does not and cannot independently obtain evidence for the expert’s claim, which may even be beyond his comprehension. Even if he is himself a specialist in the same field, he cannot, in his lifetime, conduct inquiry that will personally provide him with direct evidence for most of the beliefs he has acquired from testimony (Hardwig 1985,

339‐348).9

Based on these observations, Hardwig (1985) makes an argument, which I reconstruct as follows:

(1) Justification is required for knowledge.

(2) Evidence is required for justification.

(3) Individuals do not and cannot possess evidence for most of their

beliefs. Rather, they must trust the testimony of others.

9 Hardwig draws on an example from modern physics, in which one research paper was and could only have been written by a group of about a hundred scientists (1991, 696). Such reliance on trust, however, is not a new phenomenon. It arguably began as early as human beings started to systematically study the world. One example, from the eight century, is of natural philosopher the Venerable Bede, who, in order to develop a theory of the tides, used a network of monks all over the coast of England to systematically record data about the tides (Stevens 1985, 27). As no one individual, especially in the Early Middle Ages, could obtain such data on his own, this example shows that irreducible trust is not unique to modern science. I thank Gwyndaf Garbutt for this example.

‐ 23 ‐

(4) Either an individual can know ‘vicariously’ – i.e. without

personally possessing evidence for what she knows or perhaps

without understanding what she knows, or only a community can

collectively know.

(5) An individual cannot know ‘vicariously’.10

(6) Therefore, only the community can know.

As we can see from premise (2), Hardwig’s argument is framed within an evidentialist theory of justification. The central thesis of evidentialism is that the epistemic justification for a person’s belief that p at time t strongly supervenes on the evidence that that person has at that time (Conee & Feldman 2004, 101). It is under this assumption that Hardwig argues that qua individual a person cannot have sufficient evidence required for justification required for knowledge for most of her beliefs. The replies to Hardwig’s argument accept (2), but try to refute (3) and (5). Therefore, in this chapter, with the exception of the last section, I will assume an evidentialist theory of justification.

In light of this, let me explicitly formulate the communalist conception of knowledge I defend in this chapter (apart from the last section). Traditional analyses of knowledge are of the form ‘S knows that p’, where S is an individual subject and p is a proposition. By contrast, the communalist conception of knowledge that I defend in this chapter is of the form:

(7) We know that p.

10 In his (1985) paper Hardwig vacillates between the possibility of individuals knowing ‘vicariously’ and that of group knowledge, and seems to reject the former. In his (1991) paper he seems more favourable to this option, though he emphasizes the existence of trust as a prior and irreducible condition for justification. In his (1994) paper he seems to lean toward communalism. In this chapter, I will commit to this premise.

‐ 24 ‐

It further consists of the following two claims:

(8) Communalist Evidentialism: The epistemic justification for a

person’s belief that p at time t strongly supervenes on the

evidence that person and/or other persons in his epistemic

community possess at that time.

(9) For a normal adult individual, the justification for most of his

beliefs that are ordinarily called ‘knowledge’ does in fact

supervene mostly on evidence possessed by other people in his

epistemic community.

How are we to interpret this account? Suppose a physics student reads and understands a claim in a physics textbook. On an individualist account of knowledge we would say that she now knows the claim she has read. On a communal account of knowledge, this student does not know the physics she has read, as she does not personally possess direct evidence for what she has read. Rather, she has entered an epistemic community that collectively knows it. On this interpretation, her claim ‘I know that p’ should be interpreted as shorthand for ‘I am a member of an epistemic community that collectively knows that p’.

Let me stress that I do not claim that individual knowledge is conceptually or practically impossible. I do not deny that it is possible for a person to be individually justified in some of his beliefs. For example, I may know that I am in pain. In the presence of a basketball, I may know that there is a bouncing orange object in front of me. Rather, I argue that most of my ordinary beliefs, such that water is H2O and that

Papua New Guinea exists (if I have never been there) are not of this kind. Since our

‐ 25 ‐

conceptions of knowledge should apply to such ordinary beliefs, individualistic conceptions will not do.

Moreover, I do not draw a principled distinction between testimonial knowledge and perceptual and inferential knowledge, where the former is communal and the latter is individual. In other words, I believe that it is possible to be individually justified in holding a testimonial belief and communally justified in holding a perceptual or inferential belief. Consider the following examples. Suppose that my mother is in another room, she looks out the window and tells me that there is a blue bird outside.

Suppose that I have excellent reasons to trust her. Many times in the past I personally verified that she gives extremely accurate reports about animals and their colours.

Moreover, I know her very well, and I am sure she would never lie to me. In such a case it seems that I am personally justified in believing her testimony.11

On the other hand, suppose that a scientist is looking through the microscope at tissue sample in order to determine whether it is diseased or healthy.12 Suppose that based on her observation, the scientist forms the correct belief that the tissue is diseased. If my argument in this chapter is correct, she is not individually justified in this belief. This is because she has not personally studied the differences in observed visual patterns between healthy and diseased tissues. She is relying on other researchers, whom she may not know personally, who have carried out the relevant inquiry and published their findings in scholarly journals. Moreover, she does not personally possess evidence to the effect that her microscope reliably shows an accurate image. She is relying on instruments builders, whom again she may not know

11 See also footnote 15 below. 12 I borrow this example from Douglas (2000).

‐ 26 ‐

personally, to have tested the microscope. Therefore, although her belief is based on her own observation and inference, she is not individually justified in holding it, or so I argue.

Second, my conception is neutral on the ontology of groups, group agency and group belief. I deny Schmitt’s claim that the view that justified belief is social entails that

‘groups exist, that they have beliefs, and that these beliefs are justified’ (1994, 257). To understand why, let me draw a parallel from the philosophical discussion of group rights.13 There are two competing conceptions of group rights. Under the first conception, which is distinguished by the agent who holds the right, a right may be a group right if only the group, acting through its leadership, has the power to invoke or waive the right. This conception is problematic, because in order to establish a group right we have to show that there exists a collective social whole that is irreducible to its members in the sense that its welfare is independent from the welfare of each of its members and that this collective has the agency to act (Green 1991, 319‐20; Pinto 2009,

34‐6). The parallel epistemic case of group belief will be a belief that is held by a group irreducibly to most or all of its individual members.14

The second conception of group rights, which is distinguished by the good the right protects, avoids the problem with the first conception. Under this conception, initially proposed by Raz (1984, 189‐90) and further developed by Réaume (1988), a

13 I thank Meital Pinto for bringing this literature to my attention. 14 Intuitive examples of collective beliefs may be found instatements such as ‘the search committee believes John is the right candidate for the job’ or ‘the CIA knew that the 9/11 attacks were going to happen, even though no member of the CIA believed so’. There are two competing type of theories of group beliefs, non‐summative and summative. For non‐summative accounts, see Gilbert (1987), Searle (1990/2002) and Tuomela (1992); for a summative account see Corlett (1996, 83‐90). For a discussion of questions of group belief and agency that emphasizes their relevance to social science research, see Adler (2005, 22‐27). For a discussion of the properties groups need to have to hold collective beliefs, see Wray (2007); for an account of social cognition without group agency, see Giere (2007).

‐ 27 ‐

right is a group right if it protects a participatory good, such as language, which can only be produced or enjoyed by many. Because this conception focuses on the good the right projects rather than the agent that holds it, it is consistent with methodological individualism and need not make controversial ontological claims about groups.

Similarly, my conception of epistemic communalism need not commit to ontological holism about groups or group beliefs. Rather, it holds that knowledge is communal because the justification for most of our knowledge is a participatory good, which can only be produced by many.

Last, let me distinguish the debate about epistemic individualism and communalism from the debate about reductionism and non‐reductionism in the epistemology of testimony. Making a case for reductionism, Hume argues that in order for a person to be justified in believing a testimony, he needs to be able to reduce it to his knowledge from perception and reasoning. For example, he writes that an Indian prince who has never seen water freeze is justified in disbelieving a European’s testimony about snow, as he lacks the necessary experience for making justified inductive inference on this matter (Hume 1748/1988, 104).15 By contrast, Hume’s contemporary non‐reductionist Reid argues that because the propensity to speak the truth is part of human nature, testimony is a source of knowledge on a par with senses and reasoning ( 1764 /1997 , 193). Contemporary philosophers still debate this issue.

15 Hume qualifies this claim by saying that the prince may be justified if the testimony is ‘very strong’. This means, I suppose, that if, for example, the person who tells him about snow has so far been extremely reliable on matters with which the prince has first‐hand experience, the prince may make a second‐order inductive inference about the testifier’s reliability and reason that he speaks the truth in this case as well. Such justification is possible because the freezing of water does not contradict the prince’s already established knowledge about the laws of nature – he simply does not know what happens to water in cold climates. In cases where a piece testimony contradicts such prior knowledge, such as testimony about the occurrence of a miracle, Hume argues that the hearer can by no means be justified in believing it.

‐ 28 ‐

Coady (1994) argues that reductionism fails because Hume’s requirement is too demanding: A person’s personal experience is too narrow to justify his vast testimonial beliefs; but Fricker (1995) argues that a less demanding reductionist requirement that a person critically assess the content of a testimony and the speaker’s sincerity and competence before trusting the testimony is still in order.

Contemporary reductionists and non‐reductionists agree that when a hearer possesses defeaters, she does not have an epistemic right to trust the speaker’s testimony. The bone of contention between them is whether a hearer has a presumptive epistemic right, derivable perhaps from some high epistemic principle, to believe a speaker when she has no positive or negative reasons to trust him. Non‐reductionists say he does; reductionists say he does not. Hardwig is often wrongly interpreted as defending non‐reductionism. In fact, however, his account is consistent with both views.

He grants that a hearer can critically assess, at least to some extent, an expert’s sincerity and competence (Hardwig 1985, 342). He does not specify whether such assessment is required for the hearer to trust the speaker. He rightly notes that ordinarily, cases where there are no positive or negative reasons to trust the speaker are rare. 16

The communalists’ main claim, which individualists dispute, is that whatever grants a hearer’s epistemic right to trust a speaker’s testimony does not amount to sufficient epistemic justification required for knowledge. For communalists, readily available reasons to trust or distrust a speaker typically do not constitute sufficient

16 In an attempt to refute reductionism, Lackey constructs a thought experiment involving a testimony from an extraterrestrial alien’s diary, in which, so she argues, the recipient of the testimony has no positive or negative reasons whatsoever to trust or dist rust the testimony, but does nevertheless not acquire knowledge from the testimony (Lackey 2 008, 168‐175). It is exactly because in ordinary settings we are swamped with prima facie reasons to trust or distrust the testimonies that Lackey needs to go to such fanciful extremes.

‐ 29 ‐

evidence, if any, for the truth of the testimony. Such evidence, they claim, is usually only jointly possessed by a community. In the next section I defend this claim, and in section

3, I argue that even if we grant that reasons to trust a speaker constitute such sufficient evidence, individuals cannot possess it on their own.

2. Three Objections to Epistemic Communalism

2.1. The Objection from Convergence of Multiple Confirmations

In his reply to Hardwig, Adler argues that a person can possess sufficient evidence required for justification by gaining multiple testimonial confirmations of the same report, as they tend to converge on the truth in the long run (Adler 1994, 266).17 It is not clear, however, how Adler’s objection counters Hardwig’s claim about the irreducible role of trust in knowledge. Such subsequent multiple confirmations are testimonial and are as dependant on trust as the first testimony. They are the same in kind as the original testimony – they are merely more of the same. How can they confer justification on individual knowers if the original testimony could not?

One may argue that although subsequent confirmations do not differ from the first testimony in kind, they differ in degree. It is the number of confirmations that makes an epistemic difference. Arguably, once a testimony has been confirmed by a critical mass of subsequent reports, a person is justified in believing it. There are, however, difficulties with this response. First, numbers themselves do not make a difference. If all of the confirmations I have received originated from the same source, their number is meaningless. Second, we often encounter multiply confirmed yet false testimonies, and it seems impossible to draw a principled distinction between them and

17 Adler mentions additional ways individuals can independently confirm the testimonies they hear. The arguments for these other ways are more developed in Fricker (2002) and Goldman (1999; 2001), which I discuss sections 2.2 and 2.3, respectively.

‐ 30 ‐

multiply confirmed true reports. Webb (2004) demonstrates this claim with urban legends – ubiquitous false stories that are told as true stories that happened to somebody real. He examines several proposed principles that aim to single out urban legends from true reports, such as ‘never trust a story about a friend of a friend’ or

‘never trust when the story is “too good a story”’, and shows that they are all too permissive or too restrictive. Rumours and false myths are other kinds of such reports.18 If no principled distinction between true and false reports can be drawn, the mere fact a report is multiply confirmed does not justify our belief in it.

Empirical research also militates against the convergence‐on‐truth thesis.

Lewenstein (1995) tracks various reports (testimonies, emails, faxes, news stories, etc.) physicists received during the 1989 Cold Fusion affair. He argues that the numbers of confirming and disconfirming reports of the same claim fluctuates with time. At some point, some claim seems to be confirmed and later it does not. He argues that in the long run successive reports do not always converge on the truth. For example, successive reports may not converge on the truth because people lose interest in the story and stop talking about it with other people (Lewenstein 1995, 424).

Last, as Goldberg (2010, 154‐84) argues, in cases where multiple reports do reliably converge on the truth, this is so because there are communal epistemic mechanisms in operation. When someone says ‘if that were true I would have heard

18 During a conversation with a fellow graduate student, I was surprised when he incidentally mentioned as a fact that Hitler was of Jewish descent. When I asked him how he knew that, he said that this is something his high school history teacher had taught him. As it turns out, this almost surely false claim, which is sometimes taught as a fact in North American schools, is based on a speculative theory about Hitler’s ancestry. It is not the only false report which is often mentioned as a fact in classrooms. Having been a teaching assistant in history of science courses, I am struck by difficulty to eradicate common false myths about the history of science, against which there is conclusive historical evidence. Such false myths are still commonly mentioned as facts in science courses, and persistently appear in students’ exam answers in history of science courses, even after they have been taught the truth.

‐ 31 ‐

about it by now’, she relies on the fact, which she cannot verify for herself, that other people in her epistemic community have adequately carried out the necessary inquiry, and that members of her epistemic community are reliable sharers and distributors of information. Thus, inasmuch as multiple confirmations do tend to converge on the truth in the long run, this is just more grist to my communalist mill.

2.2. The Objection from Indirect Social Evidence

One of Hardwig’s main claims is that a hearer’s reasons to trust or distrust a speaker’s testimony that p do not constitute evidence for him for p. This ‘no‐evidence’ thesis has attracted criticism, according to which such reasons do constitute evidence – testimonial evidence, and hence the hearer can be justified (Almassi 2007; Schmitt

1988). The ‘no‐evidence’ thesis, however, is a red herring. If one insists, good reasons to trust a speaker may be called ‘evidence’, but this is merely a verbal manoeuvre

(Hardwig 1988, 311). The question is substantive rather than linguistic. Whether we call them ‘evidence’ or not, the question is whether an individual’s reasons to trust a speaker asserting that p amount to sufficient justification for believing that p. I argue that they do not.

Fricker argues that such reasons constitute sufficient evidence that amounts to justification required for knowledge, hence epistemic communalism should be resisted

(2002, 374). Fricker focuses on the problem of how individual scientists may justify the trust they give each other’s testimonies. The gist of her proposal is that scientists have plenty of evidence about their peers’ competence and trustworthiness. Such evidence is based on scientists’ personal acquaintance with their peers, knowledge of their role and

‐ 32 ‐

status in society, and knowledge of their commitment to norms of competence and truth‐telling (2002, 382‐3).

It is not clear, however, in what sense for Fricker scientists are justified qua individuals. If communal norms justify individual members’ beliefs, then justification resides at the level of the community rather than the individual. The shared norms of the community are doing the justificatory work. It is not clear in what sense these norms apply to individuals rather than the community, as they represent a joint commitment of the community. Any substantive analysis of them needs to be done at the level of the community rather than its individual members (cf. Longino 1990, 62‐82;

2002, 128‐135).

More importantly, Fricker’s claim leads to scepticism of the kind that is referred to as ‘contingent real‐world scepticism’ (Feldman 2006, 217) or ‘live scepticism’

(Frances, 2005), namely real doubts about the status of our beliefs as opposed to more familiar sceptical doubts that arise as part of a philosophical exercise. Fricker assumes that scientists’ appearing to be trustworthy is correlated with them actually being so.

But how can a person be individually justified in believing that the two are indeed correlated? How can a scientist know that her colleagues who seem respectable and present themselves as reliable and trustworthy are indeed reliable and trustworthy? To do that, she needs to verify on her own at least some of their claims to establish their reliability. It is exactly her difficulty to do so that motivates the communalist account.

Fricker’s account cannot rule out the picture depicted by contemporary sociologists of science, who emphasize scientists’ over‐reliance on social trust indicators rather than substantive evidence. For example, Sir Arthur Eddington’s

‐ 33 ‐

famous 1919 claim to have confirmed Einstein’s General Relativity was based on poor and inconclusive experimental evidence, and was trusted by his colleagues and the general public because of his reputation (Collins & Pinch 1993, 43‐54).

Attempting to diffuse this example as a support for Hardwig’s case, Almassi

(2009b) argues that Eddington’s British colleagues in astronomy and physics had the required expertise to evaluate his claim on their own; hence their acceptance of his testimony was not based on blind trust. This response, however, misses Hardwig’s point. Even if they could verify Eddington’s claim they could not, in their lifetime, verify every scientific claim they encountered. They, just like contemporary scientists, had to rely on social proxies such as an experimenter’s reputation to decide which testimonies to trust. They could not personally gain sufficient evidence that social proxies such as reputation are generally correlated with reliability.

2.3. The Objection from Reliability Indicators

If indirect social evidence cannot justify individuals’ beliefs in expert testimonies, is there other evidence that can? Goldman thinks there is. He argues that while laypersons cannot usually evaluate the evidence on which experts base their claims, they can still obtain justification for believing them, as some experts’ statements are independently verifiable by laypersons or may become independently verifiable over time.19 For

19 Goldman writes: ‘The critical point Hardwig ignores is the possibility of knowing that a specialist is an expert without knowing how or why he is an expert, just as it is possible to know that an instrument or piece of equipment is reliable without knowing how it works’ (1999, 271). In addition my argument against the first part of Goldman’s claim, there are also doubts whether the second part about instrument reliability is correct. In the sciences, establishing the reliability and accuracy of measuring instruments is a complicated process carried out by metrologists – calibration specialists who work outside the research lab at commercial laboratories or national bureaus such as the U.S. National Institute of Standards and Technology (NIST). Metrologists have excellent knowledge of how their instruments work in the form of detailed theoretical models. Research scientists can afford ignorance about the inner working of their equipment only because they trust metrologists to do their work properly and have the relevant evidence. I thank Eran Tal for this comment.

‐ 34 ‐

example, when an astronomer predicts the occurrence of a future eclipse, laypersons cannot evaluate his prediction by themselves, but once the eclipse happens, it becomes verifiable by laypersons. Such statements may be used as reliability indicators. By following an expert’s truth track record, a layperson can establish the expert’s reliability and be justified in forming beliefs based on her testimony (Goldman 1999, 271‐2; 2001,

106‐7).

An example of such reliability indicators is weather forecasts. While a layperson cannot evaluate the reasons behind a forecaster’s prediction of rain, he can look out the window and see if it rains (Goldman 1999, 79‐82). Other examples are successful medical treatments and mechanical repairs. It is not by accident, Goldman argues, that physicians and car mechanics are successful in treating people and fixing cars. Their success is due to their relevant knowledge. Their success records give laypersons justification to believe their testimonies (1999, 270‐1; 2001, 107‐8).

How promising is this as a response to communalism? As Goldman admits, it is often hard to find uncontroversial reliability indicators, especially in the frontiers of science (1999, 271 n10). In complex scientific matters, there are few, if any, statements the correctness of which a non‐specialist can evaluate.

Consider this example. Nuchal translucency and nasal bone ultrasound screenings are two types of scans performed in the first trimester of pregnancy. They aim to determine a foetus’ chance of having a chromosomal disorder. An ultrasound image of the foetus is taken, and the chances of its having a chromosomal disorder are calculated based on dimensions of the foetus’ nuchal fluid or nasal bone. The reliability of the outcome depends inter alia on the specialists’ proficiency, which is manifested in

‐ 35 ‐

her taking a good image, correctly interpreting it, correctly calculating the chances based on certain parameters and tables, etc. (Rosen et al. 2007; Sheppard & Platt 2007).

Prospective parents who want to choose a specialist to perform these tests usually have no epistemic access to the methods she uses,20 and have no reliability indicators.

Eventually having a healthy baby is not a reliability indicator, as the baby has a low chance to begin with of having a disorder. The unfortunate case of having a sick baby who was not detected is also not a reliability indicator, because these scans give only a probability and have a certain rate of false negatives. The parents cannot know if they unfortunately fell on the wrong side of the statistics, or if the disorder would have been identified had they turned to a different specialist. As much as we hate to admit, we often choose specialists based on how nice, reassuring or authoritative they are, but there is little reason to think that such personality traits are correlated with reliability.

In science and medicine, such lack of reliability indicators seems the rule rather than the exception.

Goldman also overplays the salience of reliability indicators. Were the pills the doctor gave me really necessary, or would I have overcome my cold anyway? Was the flu shot really the reason I did not contract the flu? Was the motherboard of my computer really defective, as the technician had claimed, or was the problem more minor? Did my car break down because the battery died and had to be replaced, as the mechanic had claimed, or would it have been enough to recharge it? Did Brazil’s economy recover because its government followed the World Bank’s guidelines or

20 The lack of epistemic accessibility to the specialist’s methods raises an ethical problem, as under the doctrine of informed consent, the pregnant women undergoing the scan must understand the information she have been given. However, studies show that many women do not understand basic facts about the purpose of the scans or comprehend the meaning of the probabilities involved in it (Seavilleklein 2009, 70‐1).

‐ 36 ‐

despite that? There is more than one way to fix a car, treat a patient, or help a failing economy, and the success of the treatment is often neither an indicator of the correctness of the diagnosis nor the reliability of the expert.21 Good and conclusive reliability indicators are too scarce to justify the many expert testimonial belief a normal adult has.

2.4. Conclusion

In this section, I have argued that reliability indicators, indirect social evidence and multiple testimonial confirmations cannot carry the evidential weight of justifying trust in expert or other testimonies. But even if we grant that they can carry this weight, there is a deeper problem with them. As I argue in the next section, an individual cannot obtain any such evidence in sufficient quality and quantity by herself, hence epistemic individualism fails.

3. The Argument from Epistemic Impossibility for Epistemic Communalism

Hardwig’s argument for communalism is that individuals do not possess justification in the form of personally obtained direct evidence for most of their true beliefs – only a community jointly possesses such justification, hence knowledge is communal. To resist this conclusion, both Fricker and Goldman try to show that personally obtained direct evidence about the content of the testimony is not necessary for individual justification, and argue that other reasons to trust experts, such as indirect social evidence about the sources’ sincerity and competence or external reliability indicators, can also constitute such justification.

21 This argument resonates with anti‐realist arguments in the philosophy of science that doubt whether one is licensed to infer the truth of a scientific theory from its empirical success. See, for example, van Fraassen (1980, Ch. 1) and Laudan (1981). For a response see Psillos (1999).

‐ 37 ‐

So far I have argued that such indirect evidence does not amount to adequate justification required for knowledge. In this section I argue that even if we grant

Fricker’s and Goldman’s claims, individuals cannot possess such indirect evidence in sufficient quantity and quality to amount to justification by themselves. Only a community can collectively possess such indirect evidence, and therefore individualism fails.

Take Goldman’s example of the weather forecaster. According to Goldman’s line of reasoning, a layperson cannot evaluate the reasons on which a forecaster bases her predictions, but he can verify if these predictions come true. These first‐hand verifications constitute his evidence for the belief that the forecaster is reliable. The layperson, however, can only be at a certain place at a certain time. He can only look out his own window, and can only know if the forecaster’s predictions are true with respect his street on a particular day. The forecaster may be wrong about the weather on that street on that particular day, but still be generally reliable, or happen to be right that particular day, but be generally unreliable. The indicators the layperson possesses about the forecaster’s reliability are anecdotal, partial and lacking. They cannot justify the rest of the beliefs he forms based on her testimony.

Can the layperson improve his situation? He can call his friends who live in other cities and ask them if the weather predictions have come true in their cities. But the knowledge he gains from his friends is testimonial. He must assume his friends are competent and honest. Moreover, his friends’ testimonies are themselves insufficient for establishing the forecaster’s reliability. Their testimony is only about what is happening outside their window – this is still anecdotal and unsystematic evidence. In

‐ 38 ‐

order to really establish the forecaster’s reliability, there must be a network of people, who systematically record and report the weather over time and compare their reports with her predictions. For this to work, they must trust each other. Only by being part of this community, can the layperson know if the forecaster’s predictions are, for example, better than chance. He cannot obtain enough reliability indicators on his own.

One can argue that the layperson need not be in a community of people to obtain sufficient reliability indicators. Rather, he can install many webcams in different locations, watch them online on his computer screen, systematically record the weather in these places over time and obtain enough reliability indicators. This objection, however, hardly impresses the communalist. Sure, if he had enough money, time and motivation – something which most people lack, he could in principle obtain enough reliability indicators by sitting all day in front of the computer screen watching webcams. But if he spends most of his time doing that, he will not have time to obtain evidence for all of his other beliefs – he will merely have very accurate beliefs about the weather. Recall that communalists want our concept of knowledge to adequately account for our actual beliefs, especially scientific and expertise‐based beliefs.

Therefore communalists are not impressed by what people can do in principle in order to personally obtain justification for their beliefs, but with what they can realistically do in practice.

Let us then turn to a more realistic example. Suppose Bob suffers from depression and he turns to Alice, his family physician, who prescribes an antidepressant. Suppose Bob wants to know if this antidepressant is efficacious. (I am assuming Bob is interested in knowledge and not just in overcoming his depression.)

‐ 39 ‐

Bob can start taking the antidepressant, wait and see for himself whether it works, or so

Goldman would argue. But if Bob starts to feel better, this may be merely a placebo effect; and if he does not start to feel better, it may just be that the antidepressant is not efficacious for him, but still works for most people. Bob’s personal experience is not enough to go on for knowing if the antidepressant really works. Only a systematic clinical trial can find out whether the antidepressant is efficacious. This is not as bold a claim as it may seem to some individualistically‐minded epistemologists – this is common scientific practice and exactly why the regulator requires that clinical trials be done in order to establish the efficacy of a drug.

Does Bob have other ways to obtain justified belief on this matter? Maybe he can ask his physician Alice, assuming he has good reasons to trust her. In such a case, or so

Fricker would argue, Bob can be justified in believing Alice’s testimony. But Alice is similarly not individually justified in her beliefs about the efficacy of the antidepressant.

Her beliefs are largely based on the reports she reads in professional journals describing relevant clinical trials. Alice’s knowledge of the efficacy of the drug is largely testimonial, and its quality depends on the trust all the people involved in these trials deserve – the patients accurately reporting their condition, the clinicians competently and impartially processing the data, etc. As an individual, each person possesses only a segment of the relevant evidence, and only collectively do they possess it all. I do not argue that a physician’s personal clinical experience plays no role in justifying her beliefs about the efficacy of drugs or that clinical trials necessarily override her personal experience. My point is that by herself she cannot possess systematic enough data to justify her beliefs on this matter.

‐ 40 ‐

I have used the antidepressant example for a reason. Concerns have been raised about the believed efficacy of some Selective Serotonin Reuptake Inhibitor (SSRI) antidepressants. Pharmaceutical companies have been suspect of giving clinicians economic incentives to skew results and of selectively publishing only positive trials, while repressing negative trials. Such conduct has been discussed in the medical literature in explicit terms of abuse of trust (Brown 2008, 193‐4). Due to the commercialization and corruption of research, senior scientists believe much of the clinical medical research may no longer be trusted (Angell 2009). It is exactly because there is no one individual who knows that the drug is efficacious and that obtaining such knowledge depends on irreducible trust that such sceptical worries arise. While such concerns have arisen in the medical context, the general worry applies in principle to any collaborative research which is responsible for the vast majority of the beliefs we hold and call ‘knowledge’ (cf. Grinnel 1999).

I do not argue for blanket scepticism about our scientific and medical knowledge. There are ways to improve our knowledge such as better regulation, better research protocols, and disincentives for clinicians to skew trials. My point is that such scepticism cannot be addressed within an individualistic conception of knowledge. If we ask ‘when does an individual justifiably believe that p?’ the answer will almost always be ‘never’. Rather, especially dealing with real‐world sceptical worries, we should ask

‘when do we justifiably believe that p?’.

4. Reliabilism to the Rescue of Individualism?

So far I have framed my argument about the communal nature of epistemic justification in terms of evidence. So understood, the question has been whether a person’s reasons

‐ 41 ‐

to trust another person’s testimony constitute sufficient evidence, if any, for justifiably believing the content of her testimony. One may thus think that the debate between individualists and communalists is an in‐house debate between evidentialists that need not concern reliabilists or externalists in general. In this section I argue that this is not so.

Some externalist theories of knowledge conceptualize knowledge in terms of counterfactuals an individual’s true belief must meet. Such ac counts define knowledge as true belief that satisfies certain modal conditions, such as being ‘sensitive’ and perhaps also ‘safe’. 22 Such analyses of knowledge dispense with the notion of justification altogether. They are, however, merely formal conceptualizations of knowledge. They do not make the notion of justification redundant, as justification is the reason a belief meets these counterfactuals. The fact that it is possible to give a formal analysis of knowledge without invoking the notion of justification does not mean that agents’ beliefs can be knowledge without being justified. A need for a theory of justification still arises if we are interested in answering the substantive question of why certain beliefs, and not others, are knowledge.

Externalists about justification deny that an agent’s mental state is what ultimately justifies any belief she has (Pappas 2005). The leading externalist theory of epistemic justification is reliabilism. For reliabilists about justification, what determine if a subject’s belief is justified are those things in the world that are causally responsible for the fact that the belief‐forming subject ends up with a true rather than false belief.

22 A standard formulation of these conditions is the following. Sensitivity – An agent S has a sensitive belief in a true contingent proposition p =df in the nearest possible worlds in which p is not true, S no longer believes p (Pritchard 2008, 438). Safety – An agent S has a safe belief in a true contingent proposition p =df in most near‐by possible worlds in which S believes p, p is true (Pritchard 2008, 446).

‐ 42 ‐

Reliabilism comes in many flavours. According to process reliabilism – the leading reliabilist theory of justification, a belief is justified if it is formed by a reliable process, i.e. a process that generally generates more true beliefs than false ones (Goldman 2009).

Drawing on the conclusions I have reached in the previous sections, I argue that just like evidentialists must concede that the evidence on which the justification for most of a normal adult’s knowledge supervenes is jointly dispersed among other members of her epistemic community, so do reliabilists, process reliabilists included, must concede that the truth‐yielding factors which they regard as responsible for the justification of most of an adult individual agent’s beliefs are similarly jointly dispersed among other members of her epistemic community, rather than being solely or mostly confined to that individual agent.

My quarrel is not with externalism or reliabilism per se. Rather, it is with the individualistic mindset that dominates orthodox reliabilism. As Goldberg (2010, 44) observes, orthodox reliabilists assume that the processes through which beliefs are formed never extend beyond the boundaries of the individual believer. Goldberg argues that this is assumption is mistaken and the notion of a reliable process should extend beyond the individual believer to his social environment.

As Goldberg (2007, 209‐26) argues, the individualistic assumption inhibits the explanatory abilities of process reliabilism. For instance, individualistically minded process‐reliabilists cannot explain how young children, whose cognitive processes are not yet reliable and who cannot reliably discern their parents’ true testimonies from others’ false testimonies, can nonetheless acquire knowledge from testimony. He argues that cognitively immature children are reliable consumers of testimony only relative to

‐ 43 ‐

the epistemically secure environment their parents generate for them. In such an environment they shield them as much as possible from false testimonies and encourage them to develop a critical attitude toward unreliable sources, like other children and deceiving uncles who enjoy pulling their legs. The general lesson for

Goldberg is that what justifies a subject’s belief is not the reliability of the subject’s cognitive belief forming process. Rather, it is the de facto reliability of the subject’s sources of belief, and the social conditions that make that subject a reliable consumer of knowledge in a given social environment.

I suggest that the relations Goldberg describes between young children and their parents are analogous to the relations between laypersons and experts. I argue, based on the discussion in the previous sections, that like children who cannot reliably discern between true and false testimonies without their parents’ help, laypersons cannot reliably discern by themselves between true and false testimonies within a given domain of expertise without the help of experts. Similarly to parents, who create for their children social conditions in which they can be reliable consumers of testimony, so do experts (at least try to) create social conditions, through mechanism such as peer review and drug regulation, that allow laypersons to be reliable consumers of expert testimony. Similarly to parents who guide their children and help them become critical consumers of knowledge as adults, so do experts train laypersons, such as undergraduate students, to become experts themselves. In the same way the justification of children’s beliefs is determined by the de facto reliability of their parents and the surrounding social conditions, so is the laypersons’ justification determined by

‐ 44 ‐

the de facto reliability (or unreliability) of experts and the social context in which they are situated.23

To understand why this is so, consider the following example. Goldman describes a case in which a novice bird‐watcher and an expert see a bird. They both spontaneously form the correct belief that this is a pink‐spotted flycatcher, but only the expert is justified, as the novice is jumping to a conclusion out of excitement (Goldman

2008, 67). Goldman argues that process reliabilism explains why the expert’s spontaneous belief is justified, but not the novice’s. The cognitive process employed by the expert reliably matches between the bird visual experience and his memories of pink‐spotted flycatchers. The novice just guesses. Thus, the expert’s belief forming method is reliable, while the novice’s is not (Goldman 2008, 75‐6).

Let us change the story such that instead of jumping to a conclusion, the novice asks the expert what kind of bird this is. The expert replies truthfully, and the novice forms the respective true belief. Suppose for the sake of the argument that the novice is justified in his belief. What would explain his being justified? Orthodox process reliabilism will explain the novice’s justification in terms of the reliability of his testimonial belief forming process. But this is surely wrong. The reliability of the novice’s cognitive process has little to do with the justificatory status of his belief.24

23 Analogies are never perfect. By drawing an analogy between parents and their children and experts and laypersons, I am not suggesting an authoritative concept of knowledge in which laypersons blindly accept expert testimony. In addition, as I argue in Chapter 2, I reject the so‐called deficit model of scientific communication, in which scientific communication is a one‐way channel, where the lay public passively assimilates scientists’ claims. I am not denying the possibility or desirability of public engagement with science and the public’s ability to affe ct scientific epistemic standards. On the contrary, my conceptualization of the knowing subject as community that collectively possesses epistemic justification is consonant with the public being involved and active in the process of knowledge generation, bearing in mind that scientists also occupy the layperson position when they rely on other experts and non‐experts for having justified beliefs. 24 This debate runs parallel to the debate about the credit theory of knowledge, which roughly states

‐ 45 ‐

Suppose the expert, like a deceiving uncle, is usually in conning mood. Suppose that he indented, as usual, to pull the novice’s leg, but decided on the spur of the moment to reply truthfully just this once. Had the expert not changed his mind at the last moment, the novice would have formed a false belief. The true belief he did form is unjustified, as it was just accidentally true.

Both in the case of an honest expert and a conning expert, the novice forms his belief using the same testimonial belief forming process. Moreover, both the honest and conning experts form their respective beliefs using the same (reliable) belief forming process. Therefore, the reliability of the novice’s testimonial belief forming process is not responsible for the justificatory status of his belief, and the reliability of the experts’ relevant processes is only partly responsible for it. What really matters for justification is the de facto reliability of the expert.

Suppose that an individualistically‐minded process reliabilist wanted to insist on the role of the agent’s belief forming process. She would have to insist that the novice’s testimonial belief forming process could reliably discern between true and false expert testimonies on such occasions. How could such argument be supported? The visual experience of the bird would not play a part in this explanation, as ex hypothesi the novice cannot reliably identify the bird. The process reliabilist may argue that certain social cues in the situation might help the novice reliably discern between true and false expert testimonies. But what could they possibly be? Presumably, such social cues

that knowledge is true credit‐deserving belief. As Lackey (2007, 354‐6) argues, in typical cases of testimonially obtained beliefs, such as a stranger’s getting directions from a local person, the stranger can be easily mislead, and while he deserves some credit for using reliable faculties such as hearing and seeing to obtain the testimony in question, the reliability of these faculties is not the salient reason for the belief being true. Lackey does not consider a communalist credit theory of knowledge in which an epistemic community jointly deserves credit.

‐ 46 ‐

would be multiple confirmations, indirect social evidence, or external reliability indicators.

But as I have argued, such indications are insufficient for justifying the novice’s belief in such a case. If such data are insufficient as evidence, they are just as insufficient as input data based on which the novice’s cognitive processes may reliably discern between true and false testimonies. It does not matter whether such indications are internally accessible to the novice or not. It does not matter whether they are thought of as evidence for the novice, or as sensory input data for his belief forming process. If my argument in the previous sections is right, then the information they contain is an insufficient basis for reliably discerning between true and false expert testimonies.

In my view, externalists should not find this conclusion too hard to resist, as they regard justification as external to the knowing subject. In particular, reliabilists, who regard the causal history of a belief, which usually extends beyond the believing subject, as relevant to its justification, should not find this conclusion so disturbing. Some reliabilist theories of knowledge, such as Goldberg’s Reliabilist Social Epistemology

(2010) and Corlett’s Social Epistemic Reliabilism (1996) give up the individualistic assumption and provide social‐reliabilist analyses of justification. It would be wise for other reliabilists to move in that direction as well.

Conclusion

I have argued that the justification for most of an individual’s true beliefs is possessed by or located in the epistemic community to which she belongs. Any proper evidentialist theory of epistemic justification must acknowledge that the evidence that justifies most of a person’s beliefs is jointly possessed by other members of her

‐ 47 ‐

epistemic community. Similarly, any proper reliabilist theory of justification must acknowledge that the factors that determine the truth of a subject’s beliefs extend far beyond that subject into the epistemic community to which she belongs.

If epistemology is to be relevant to our vast set of ordinary beliefs that we usually call ‘knowledge’, especially knowledge we acquire from science, and if it to make a positive contribution in cases where we have real doubts about our beliefs, for example because of the corruption of medical research, then it should give up its individualistic dogmatism and adopt epistemic communalism instead. As the aim of this dissertation is to develop a framework we can use to assess our theories and beliefs when we have such doubts, in the rest of this dissertation I will assume a communal conception of knowledge.

‐ 48 ‐

Chapter 2

Relativism and the Limits of Social Explanation:

A PRIME Example

Introduction

In the previous chapter I argued that individualist conceptions of knowledge cannot successfully deal with sceptical doubts about our ordinary beliefs that we usually call

‘knowledge’. As I have mentioned, due to the corruption of medical research, for example, and the fact that scientists must irreducibly trust each other, sceptical doubts about the status of some of our medical knowledge have indeed arisen, and a worry has been expressed that it might be skewed to serve the interests of big pharmaceutical companies.

But what about the rest of our beliefs that we regard as knowledge? They are also for the most part dependent on irreducible trust. Maybe blanket scepticism about them is in order? Perhaps it is the case that our ordinary beliefs are not justified and are largely constructed to serves the various social interests of different social actors?

Such views are commonly associated with the sociology of scientific knowledge

(SSK). According to constructivist SSK epistemology, knowledge is socially constructed, in the sense that it can be wholly or largely reduced to social agreement between various actors. It follows from this view that it is possible to explain both stability and change in our beliefs by appealing wholly or largely to social factors such as social institutions and social interests, rather than to notions such as truth, justification and rationality (Hacking 1999, 84‐92).

‐ 49 ‐

Such views, of course, are controversial. Critical responses to social constructivism try to show, with various degrees of success, that the epistemological assumptions underpinning SSK theories are unfounded. Some of them attempt to show that these assumptions hinder them from providing adequate social explanations of science. Some such accounts deny some of the constructivists’ premises and offer a rational or realist reconstruction of construction narratives, which they argue to be more plausible (Giere & Moffatt 2003; Brown 2001, 115‐43; Goldman 1999, 225‐30).

I will also give an argument to the effect that the constructivist conception of knowledge as a mere social agreement hinders SSK explanatory success. My approach, however, will be different. In this chapter, I will accept for the sake of the argument constructivist premises and argue that constructivist explanations fail on their own terms. I will present a case study that will demonstrate that the existence of independent epistemological standards needs to be assumed in order to adequately explain its outcome.25

This case study is the following. In August 2002, three Indian computer scientists, Professor Manindra Agrawal and his two students Neeraj Kayala and Nitin

Saxena, from the Indian Institute of Technology Kanpur (IIT) published a paper,

‘PRIMES is in P’, online. It presented a ‘deterministic algorithm’ which determines in

‘polynomial time’ if a given number is a prime number (the AKS algorithm). The story

25 For other work that takes a similar approach see Silva (2005), which examines experiments in aerodynamics, and argues that discursive theory alone cannot explain the existence of a giant physical robotic model of a moth in these experiments, its role in producing knowledge and the different knowledge that would have been produced had computer simulations been used instead. This is because the theory lacks the necessary concepts to deal with the materiality of the model. See also Giere (1988, chapter 5), which based on a field study of a nuclear physics laboratory, argues that one cannot give an adequate social explanation to the physicists’ behaviour without assuming the ontological reality in which they believe.

‐ 50 ‐

was quickly picked up by The New York Times and the rest of the general press, and by this means spread through the relevant scientific communities of complexity theorists and number theorists, where it was hailed as a major theoretical breakthrough. By the time the paper was published in a peer‐reviewed journal, two years after its initial publication on the Internet, Agrawal and his students had already received wide recognition for their accomplishment.

The general media usually show little interest in theoretical developments in computer science or mathematics, important as they may be, but in this case, the general media devoted a surprising amount of attention to the story. However, the media’s interpretation of the meaning and implications of the new algorithm was very different from that of specialists in the relevant fields, who regarded the media’s interpretation as distorted.

How come a theoretical development in computer science received this much attention from the general press? In what ways was the interpretation by the press of the AKS algorithm different from that of the scientists? Did the Indian scientists have an interest in this press coverage? Can we determine which one of the interpretations of the AKS algorithm is the ‘correct’ interpretation? Why is it that the three scientists’ choice to publish their result on the Internet and in a popularized manner in the general press rather than a peer‐reviewed journal did not damage their scientific reputation among their peers?

Current sociological theory challenges the ability to clearly distinguish on independent epistemic grounds between genuine and simplified scientific knowledge, as well as between faithful simplifications and distortions. It views the demarcation

‐ 51 ‐

lines between such forms of presentation as largely contextual and unstable. It takes for granted the assumptions of SSK epistemology, according to which knowledge may be reduced to agreement that is based solely or mostly on social interests of the parties to the agreement, and try to answer questions of this sort by appealing to social factors.

I give a systematic survey and analysis of the popular press coverage of the

‘PRIMES is in P’ affair in the English language. I argue that when dealing with popularization, it is not necessarily true that distorted accounts of knowledge cannot be distinguished from faithful simplifications on independent and recognizable epistemic grounds. Additionally, the demarcation lines between distorted and non‐distorted representations of scientific knowledge are not as open to political manipulation as the new sociological view of popularization suggests.

The existence of such independent epistemic grounds will explain, in turn, the ability of Agrawal and his students to simultaneously communicate distorted and non‐ distorted accounts of their discovery to different audiences, without damaging their scientific reputations. If independent grounds for distinguishing distorted from non‐ distorted accounts did not exist, it would be very likely that Agrawal and his students’ choice to disseminate their results on the Internet and through the popular media would have damaged their reputations and negatively affected their careers. This is because not only did Agrawal and his students violate the social norms of the scientific community by turning to the popular media and communicating their research results directly to the public, these media reports were inconsistent with the consensual views among specialists. Therefore they stood in contrast to their cognitive interests. As I will show, although scientists first learned about Agrawal and his students’ achievements

‐ 52 ‐

from the media, he and his team received full recognition of their achievement from their colleagues. I will argue that this is because scientists were able to identify the

‘genuine science’ within the media’s distorted accounts.26

My case study will demonstrate that while various interests of the actors involved affect the different representations of scientific knowledge in different media, distorted simplifications of scientific knowledge are distinguishable from non‐distorted simplifications on independent epistemic grounds. Furthermore, because such independent epistemic standards exist, scientists are able to communicate different contents to different target audiences in order to promote their interests.

This chapter may also be considered as part of what Collins and Evans call a

‘third wave’ science study. I rely on current SSK theory of popularization to explain the events of the ‘PRIMES is in P’ affair (hereinafter ‘PRIMES’). However, similarly to Collins and Evans, who (hesitantly) claim that science is not entirely reducible to politics (2002,

245, 286 fn 27), this chapter calls for a reform to the current theory of popularization by acknowledging that scientific knowledge is at least partly constrained by non‐political factors, and that this fact should be used as an explanatory resource in social explanations of scientific affairs.

This chapter consists of five sections. Section 1 presents the current theory about popularization and challenges the view that distorted simplifications cannot be

26 The literature about popularization covers both the popularization of scientific knowledge and practice. With respect to scientific practice, it is argued that the media conveys to lay audiences an idealized account of the scientific method, which hides the intricate process of social conflicts and negotiations through which scientific knowledge is constructed. Because of this idealized account of the scientific method, lay people ascribe a high degree of reliability to scientists’ claims (Gregory & Miller 1998, 90‐1). Within the scope of this chapter, I only address the question of the popularization of scientific knowledge, but not of scientific practice. In other words, I address the question of what beliefs lay readers come to form from popularized reports about what scientists claim to have done. I do not address the question of the warrant that lay readers have or should have for these beliefs, although, of course, in general these two questions are not isolated.

‐ 53 ‐

clearly distinguished from non‐distorted ones. Section 2 provides scientific background from the theory of computation which is necessary for understanding the meaning and significance of the AKS algorithm in its scientific context. Section 3 systematically surveys the general press coverage of the ‘PRIMES is in P’ paper. In that section, I analyze the explicit as well as the implicit ways in which the general press gave a false impression of the implications of the AKS algorithm to lay readers. Section 4 analyzes the interests that the scientists and the press had in the media coverage of PRIMES. In that section, I identify three interests – visibility, recognition and priority – which the scientists had in the general media coverage of their algorithm. Section 5 addresses possible criticisms of my argument and discuss its methodological significance and

generalizability to other sciences.

1. Between Popularization, Simplification and Distortion

Roughly speaking, two views regarding popularization of scientific knowledge may be identified in the literature, the traditional model and the new model. They differ on three main points. First, the traditional model assumes that audiences are atomistic uninformed assimilators of information, with little or no collective internal structure

(Whitley 1985, 3). Traditionally, ‘science is the active disseminator and the fountain of meaning and agency, the public are merely the passive receivers and repositories’

(Michael 1996, 109).

Second, popularization is traditionally viewed as external to the knowledge production and validation process, which is left to non‐scientists. Scientists’ dissemination of scientific knowledge to audiences of non‐scientists is viewed as a

‐ 54 ‐

subsidiary activity that does contribute to a researcher’s reputation, or may even in fact damage it (Whitley 1985, 3).

Third, the traditional model holds an idealized notion of pure genuine scientific knowledge that it contrasts to popularized knowledge. Any differences between genuine and popularized sciences are assumed to be caused by distortion or degradation by journalists and the lay public (Hilgartner 1990, 519). Traditional communication studies therefore search for ways to improve accuracy and balance in science reporting and to avoid sensationalism and distortion (Lewenstein 1995, 407).27

The new view of popularization in the sociology of science challenges each of these three assumptions. First, the new model recognizes diversity within the public and its attitudes towards science. Sociological research shows that members of different publics construct different self‐perceptions of their interest in and knowledge of science as part of their social identity. They can also critically reflect on their own epistemological standards (Michael 1996).

In addition, the public consists of a number of readily identified audiences, some of which are important for scientific research. Some members of the public are scientists from other fields. Some belong to professional occupations, such as engineering, which claim legitimacy for their use of science. Some are university students, from which future researchers can be recruited, and some, for example policy makers, wield power to make decisions regarding scientific research. All of these types of audience treat popularized scientific knowledge differently, and to these different types of audience scientists deliver different types of knowledge (Whitley 1985, 5).

27 See Väliverronen (1993, 24‐26) for a literature review of studies associated with the traditional model.

‐ 55 ‐

Second, according to the new model, popularization has an active part in the process of producing and generating knowledge. The mechanisms are twofold. First, in order to gain general support from society and lay decision makers, scientists need to simplify scientific knowledge, popularize it and emphasize its practical value (Whitley

1985, 19). Second, affecting the view of and gaining support from outsiders, such as decision makers or the general public, may tip the scales in scientific controversies within a scientific community. Naturally, such external audiences learn and form their views of scientific knowledge from the popularized accounts that scientists provide to them.

How does popularization feed back into the scientific community? This is done by giving references to popularized sources in scholarly publications, thus legitimizing them as good science (Hilgartner 1990, 523‐24). Another way is through public or private communication between scientists and science journalists, where scientists learn from journalists and media reports about recent developments in their field before they appear in scholarly journals (Lewenstein 1995, 411‐24). The crucial point is that scientists form judgments and shape their beliefs and expectations about scientific factual claims based on popularized sources.

According to the new model, what underpins and enables these processes is the fact that scientific knowledge is produced in a social process of negotiations. As Whitley puts it:

‘Facts’ are socially constructed cognitive objects, liable to reinterpretation and change, which become established through negotiations and extensive communication among scientists. The exposition of research results to scientific audiences is a crucial component of these processes which affects what comes to constitute knowledge in that field at that time. Expository practices are not epistemologically neutral. (Whitley 1985, 11; footnote omitted)

‐ 56 ‐

This picture challenges the third assumption of the traditional model about the categorical distinction between genuine and distorted scientific knowledge. The new model rejects any notion that scientific knowledge completely transcends social context.

Since the distinction between genuine and distorted accounts assumes such a transcendental account, the new model rejects it too. As Lewenstein puts this:

… a technical paper presented at a small conference is no more ‘science’ than a multimedia extravaganza presented on an IMAX screen or at Disney World’s EPCOT Center. Both are attempts to use rhetoric to present understandings of the natural world to particular audiences. (Lewenstein 1995, 408)

According to the new model, knowledge always takes form in a particular social context. Knowledge is always tied with epistemic standards that are used to validate it.

These epistemic standards are social norms, which are always situated within a particular epistemic community and its unique way of living. Different epistemic communities may offer different yet equally valid forms of knowledge (Wynne, 1996).28

The new model holds that epistemic standards are determined by contingent matters and relative to an epistemic community or even to an individual scientist. On

Lewenstein’s view, for example, epistemic standards are in constant flux. Throughout their work, scientists happen to encounter different reports from various sources, popularized and non‐popularized, and form ideas about them in an accidental fashion.

In turn, these scientists produce other reports, which are consumed by other actors and so on:

28 Following Kusch, I take epistemic standards to be a set of exemplars (in the Kuhnian sense) that are shared by members of an epistemic community. Justifying a claim is a dialectical process in which members of the epistemic community try to show that the relations between the content of the claim and the evidence for it are similar or analogous to one of the communally endorsed exemplars. A claim counts as knowledge when the community is satisfied that this is indeed the case (Kusch 2002, 120‐30). As I will show, however, my case study militates against Kusch’s claim that epistemic standards are necessarily relative to an epistemic community, and that one epistemic community’s epistemic standards cannot be said to be objectively better than another community’s epistemic standards (Kusch 2002, Ch. 18).

‐ 57 ‐

People take in lots of information, filter it in various ways and base their judgements on a range of issues running from salience and importance through time of day and state of hunger …. Theory suggests that each reader would make a different judgement, based on completely contingent factors. No model attempting to predict the value of different types of communications can work. (Lewenstein 1995, 415; emphasis in the original)

As Broks (2006, 125) nicely puts it, Lewenstein conceptualizes scientific knowledge as part of a web in which ‘press conferences, lab reports, news programmes, emails, grant proposals, policy documents and seminars are interconnected and feeding into each other’. An alternative conceptualization is a continuum of forms of representation of scientific knowledge, in which specialist accounts are on one extreme and non‐specialist accounts are on another (Hilgartner 1991, 525‐28; Whitley 1985, 7‐

8). The important point for this chapter is that neither conceptualization invokes global or independent epistemic standards to determine whether a given account is a distortion or not. This is always determined locally based on contingent standards.

This line of analysis emphasizes the political dimension of knowledge. Both

Hilgartner and Bucchi argue that without epistemic ‘gold standards’ for evaluating scientific knowledge, the process of determining what is ‘genuine’ science and what is not becomes political. A decontextualized scientific report (if one can be imagined) would neither be genuine nor distorted. Rather, based on their interests at a given time, scientists determine whether a given account with its degree of simplification is a distortion. Scientists enjoy the exclusive social authority to demarcate ‘real science’ from ‘popularized science’, and ‘faithful simplifications’ from mere ‘distortions’.

Therefore, in order to preserve this authority in accord with their political interests, scientists can label some representations of scientific knowledge as ‘appropriate representations’ and others as ‘distorted accounts’, while blaming, for example, the

‐ 58 ‐

media for the distortion (Hilgartner 1990, 320; Bucchi 1996, 377‐8). Generally speaking, as a political resource, scientists maintain strict boundaries between science and non‐scientific forms of knowledge. Science, so they argue, is governed by superior epistemic standards of prediction and control. In their view, the superiority of scientific knowledge entitles them to social authority in the form of expertise in public matters to which scientific knowledge is relevant (Wynne 1996).

Bucchi stresses that the distinction between appropriate scientific dissemination and distorted spectacles is a political resource available to scientists. It is used, for example, to exclude colleagues or other actors from the public arena (Bucchi 1996,

387). Scientists who mainly communicate popularized accounts to the general public are typically ostracized by their colleagues, especially if they have not yet gained the reputation of serious scholars. This is because popularizers undermine scientists’ exclusive hegemony on the construction of scientific facts (Gregory & Miller 1998, 82‐

3).

On the other hand, scientists themselves learn about fields outside their own from reports in the popular media and other simplified accounts. When it suits their purposes, scientists refer to such accounts in their specialized publications, thus treating them as appropriate representations (Hilgartner 1990, 520‐24). Furthermore, when it suits their purposes, they use media reports to set epistemological standards for the closure of scientific controversies. In some cases, media reports are used as a significant resource for reaching collective agreement on what experiments would count as successful replications (Simon 2001 388‐89).

‐ 59 ‐

Under the new model, then, various representations of scientific content exist in the different media. These representations may be labelled as popularizations and used to discredit certain viewpoints and individuals in a scientific community. Alternatively, they may be labelled as appropriate representations and used to legitimate certain views and claims within a scientific community. It is because popularized content by itself is indistinguishable on independent epistemic grounds from non‐popularized content, that scientists are able to put similar representations of scientific content to opposing uses in different circumstances.

While my case study supports the new model’s view about the generative role of popularization and the active role of different types of audiences, it challenges the view that the difference between genuine knowledge, faithful simplification, and distorted simplification is purely or at least largely political. As a basis for my analysis I would like to address two points with respect to the new model.

First, I would like to challenge an implicit assumption that underpins the new model, which is that scientists enjoy an exclusive epistemological authority on the definition of science in society. In my view, when scientists operate outside their field, their knowledge is subject to different epistemic standards. Scientists often do not decide what counts as science outside their own field. Law, for example, has its own standards for distinguishing scientific from non‐scientific knowledge and ascertaining the reliability of scientific evidence.29 Even when scientists’ views about what science is do matter outside their field, they do not control the standards by which non‐scientists evaluate scientific claims. For example, scientists do not control the media’s standards

29 For a critical review of the changing legal standards for evaluating scientific expert testimony in US courts see Haack (2003), chapter 9.

‐ 60 ‐

for what is worthy of publication. Therefore if scientists want their research to appear in the general media – and I will show that many of them do have such an interest – they need to play along and present their research in the most attractive way for the media.

Scientists are not the only actors, and perhaps not the main actors, who set standards for what is presented and gets recognition as science in the public arena.

Second, as the new model rejects the distinction between genuine and distorted knowledge, it uses terms such as ‘distortion’ and ‘popularization’ interchangeably.

These words are also usually scare‐quoted, to indicate their fictitious reference. While the new model does not require a conceptual analysis of these terms, my account does. I will use the term ‘distortion’ for accounts that mislead their readers and make them form wrong beliefs. As Adler notes, distortion has an element of wilfulness or intellectual neglect, which excludes innocent misconstrual. Distorters may take advantage of existing epistemic norms within a given epistemic community, which, when applied to a report, will likely cause it to be misconstrued (Adler 2007, 383). I will use the term ‘simplification’ to denote reports that present the original account in a less detailed, less jargon‐laden or generally less complex form. Note that a simplification is not necessarily a distortion. A report can omit certain details, for example, without giving its reader a wrong impression. I will use the term ‘popularization’ mainly to denote distorted simplifications.

I suggest that scientific and non‐scientific epistemic standards are different from each other, and are relatively stable. As my example will show, scientists achieve political goals by adjusting the different accounts of knowledge they produce to these different standards, not by defining these standards and playing with them. In section 3,

‐ 61 ‐

I will give a detailed analysis of how the media distorted the meaning of the ‘PRIMES is in P’ paper. In order to do so, I will first need to give a brief technical background of the subject.

2. ‘PRIMES is in P’ – Necessary Scientific Background

As I will show in section 3, the ‘PRIMES is in P’ paper was susceptible to media distortion, as terms that are used in it, such as ‘deterministic’ and ‘probabilistic’, have different meanings and implications in the scientific and ordinary contexts. In addition, in computational theory two distinct problems exist with regard to prime numbers,

PRIMES and IFP, and the media reports have tended to confuse them. In this section, I will explain the significance of the ‘PRIMES is in P’ paper, as understood by specialists who are familiar with this problem. This background is needed to understand my argument in sections 3 and 4 of this chapter. Readers who are familiar with the theory of computation may skip to the conclusion of this section.30

2.1. Complexity Class P

In the theory of computation, problems that can be solved by computers are grouped into complexity classes according to the time it takes a computer to solve them. The time is not measured in seconds or minutes, but as a function of the length of the input; specifically, the relationship between the growth of the time function and the growth of the length of the input.

30 The following account is simplified from two leading university‐level computer science textbooks (Cormen et al 2001; Sipser 1997). Therefore this account may be considered a ‘canonical’ account. The account I give is simplified in two main ways. First, it does not include any proofs, so the claims remain without their justifications. Second, mathematical jargon and technical notat ions are largely omitted. Though simplified, my account is not distorted. My ability to give such a simplified yet non‐ distorted account is consistent with my claim in this chapter. (In section 5 of this chapter, I defend my choice to rely on this account when analyzing the media reports.) Due to my own academic background in computer science I have ‘interactional expertise’ that allows me to serve as a translator (Collins & Evans 2002, 254‐258) between computer scientists and the readers of this chapter.

‐ 62 ‐

For example, two important complexity classes are P and EXP. To class P belong all the problems that can be solved in polynomial time by a deterministic algorithm (an explanation of what a deterministic algorithm is will be given in the next subsection). In other words, the time it takes to solve problems in the complexity class P is a polynomial function31 of the length of the input.32 In contrast, to class EXP belong problems that are solved by a deterministic algorithm whose time to solution is an exponential function of the length of the input.33 The difference between these two classes is the rate growth of the functions. While polynomial functions grow relatively slowly with the growth of their input, exponential functions grow very fast. Therefore, problems that belong to complexity class P are regarded as problems which can be solved in reasonable time. In contrast, problems that belong to the complexity class EXP are considered to be problems which cannot be solved in reasonable time.

For example, let us suppose that problem K with an input of the length of x characters can be solved by a computer in x2 seconds. Because the function x2 is a polynomial, problem K belongs to complexity class P. Then, if the length of the input is

10 characters, it will take the computer 102=100 seconds – less than 2 minutes – to solve. If the length of the input is 100 characters, it will take the computer 1002=10,000 seconds – about two hours – to solve. In contrast, let us assume that problem L with an input of the length of x characters is solvable in 2x seconds. Because the function 2x is exponential, problem L belongs to complexity class EXP. Then, if the length of the input

31 n n­1 A polynomial function is a function of the form f(x) = anx + an­1x + ... + a1x + a0, where an…a0 are constants and n is a positive natural number. Examples for polynomial functions are f(x) = 3x2, f(x) = 5x27 + 12x20 + 8x4, etc. 32 More precisely, let a be an algorithm, let x be the length of the input and let the function t denote the running time of the algorithm, then a belongs to complexity class P if and only if there exists a polynomial function p such that for all x, t(x)

‐ 63 ‐

is 10 characters, it will take the computer 210=1,024 seconds – about 17 minutes – to solve it. However, if the length of the input is 100 characters, it will take the computer

2100 seconds – more than billions of years(!) – to solve it. It is important to notice that such problems are considered unsolvable, in principle, in reasonable time, regardless of the actual computing speed of contemporary computers. Even if we had computers that were, for example, 100 times faster than current computers, because of the fast growth of the running time function, if we slightly increased the length of the input, it would still take billions of years to solve problems that belong to complexity class EXP.34

2.2. Probabilistic and Deterministic Algorithms

Another distinction is made in the theory of computation between problems that have a deterministic algorithm for solving them and problems that are solved by probabilistic algorithms. If the algorithm is deterministic, then it means that it reaches the same and right answer every time it is run. In contrast, if the algorithm is probabilistic, its output depends on its ‘tossing a coin’ during its run. It does not necessarily give the same answer every time, and there is a chance that it will reach a wrong answer. This distinction has theoretical significance in the theory of computation. However, from a practical point of view, probabilistic algorithms are as reliable as deterministic algorithms. This is because we can run an algorithm as many times as we want to achieve as high a degree of confidence as we want. For example, let us suppose that an algorithm A is used to solve a problem p, and that if the correct answer to p is negative,

A has a probability of ½ for giving an incorrect positive answer. However, if the correct answer to p is positive, A will always give the correct positive answer. If we run A 100

34 For a formal definition of complexity class P see Sipser (1997) at 234‐36.

‐ 64 ‐

times, for instance, and get the same positive answer every time, there is a probability

1 of only that the answer we have is incorrect. This is an extremely low probability, 2100 lower, for example than the probability that a meteor will hit you before you finish reading this sentence.

In addition, it is important to notice that when deterministic algorithms are used in practice, they are not 100% accurate. This is because there is the possibility that there is a bug in the program which implements them, or a bug in the compiler that was used to compile them, or in the operating system used to run them, or in the hardware used to run the operating system, or that there will be an electrical fluctuation which will change the content of the memory of the computer, or that a cosmic ray will hit the computer and change the content of its memory, etc. In other words, from a practical point of view, there is no difference between the reliability of deterministic algorithms and probabilistic algorithms (Sipser 1997, 335‐6). To sum up, probabilistic algorithms are as reliable for any practical use as deterministic algorithms. Probabilistic algorithms are treated differently from deterministic algorithms mainly by theorists, who are not interested in their practical use, but in their mathematical properties.

2.3. The Difference between PRIMES and IFP

In the theory of computation, a distinction is drawn between two distinct mathematical problems. The first problem is PRIMES. PRIMES is defined as the problem of finding whether a given number is a prime number.35 (A prime number is a natural number that has only two natural number divisors, which are one and the prime number itself.)

In contrast, IFP (Integer Factorization Problem) is defined as finding the factors of a

35 A formal definition of PRIMES is: PRIMES = {n | n is a prime number}.

‐ 65 ‐

number.36 For example, given the number 6, the output of an algorithm which solves

PRIMES will be ‘not prime’, because 6 is not a prime number. In contrast, given the number 6, the output of an algorithm which solves IFP will be ‘2 and 3’, because 2 and 3 are the factors of 6 (2⋅3=6). Of course, a solution for IFP entails also a solution for

PRIMES, because if we know the factors of a number, we immediately know if it is prime or not. However, a solution for PRIMES does not entail a solution for IFP – when an algorithm that solves PRIMES gives us an answer we know whether or not the input number is prime, but we do not know its factors. PRIMES is a decision problem, namely an algorithm that solves it gives us the answer ‘yes’ or ‘no’. By contrast, IFP is a calculation problem, namely an algorithm that solves it gives us numbers that are the solution to a mathematical calculation problem.

Now, in the paper ‘PRIMES is in P’, Agrawal and his students describe an algorithm which is (1) deterministic and (2) solves PRIMES (3) in polynomial time, hence the title ‘PRIMES is in P’. Before their paper, the question whether PRIMES was in

P or not had been a long lasting open theoretical problem, hence the wide attention they got from their peers.

It is also important to emphasize that the AKS algorithm which was presented in the ‘PRIMES is in P’ paper solves the problem PRIMES and not the problem IFP. In other words, it determines if a given number is prime or not, but does not find its factors.

Currently, there is no known algorithm that solves IFP in polynomial time.

It is also important to mention that a probabilistic algorithm that solves PRIMES in polynomial time was already introduced in 1976 by Miller and improved by Rabin in

36 A formal definition of IFP is: IFP = {x | x = pq, for integers p, q>1}.

‐ 66 ‐

1981, and has been widely used since. Since, as aforementioned, probabilistic algorithms are as reliable as deterministic ones, the AKS algorithm did not have any practical implication (Sipser 1997, 339‐43).37

2.4. The RSA Encryption Algorithm

The RSA (Rivest Shamir Adelman) algorithm is an encryption algorithm that is widely used on the Internet, among other places. If two participants want to use RSA to exchange encrypted data, each of them must possess two types of key: a private key, which is known only to each of them alone, and a public key, which is known to everybody. If I want to exchange encrypted messages with someone, all I need to give that person is my public key. The public encryption key is like a key that can lock a box, but cannot unlock it once it is locked. Using my public key, the other person will be able to encrypt messages to me, but once the message is encrypted, only I will be able to decipher it, because I am the only one who knows my private key.

What is the connection between RSA and prime numbers? In RSA, part of my public key is a number n, which is the product of two large prime numbers, p and q. In order to verify that p and q are in fact prime, the probabilistic Miller‐Rabin algorithm, which solves the PRIMES problem in polynomial time, is used. As mentioned earlier, the fact that the Miller‐Rabin algorithm is probabilistic does not affect its reliability whatsoever, and therefore it does not compromise the strength of the encryption. In

RSA, my private key is computable from my public key, but not in reasonable time. If somebody other than me wants to compute my private key from my public key in order to break the encryption, he must factor n in reasonable time. In other words, he needs to

37 It is useful to note that for practical implications, the AKS algorithm runs significantly slower than the Miller‐Rabin algorithm, although both of them run asymptotically in polynomial time. Therefore, for practical implementation, the Miller‐Rabin algorithm is preferable to the AKS algorithm.

‐ 67 ‐

have a polynomial time algorithm for solving IFP. However, since currently there is no known algorithm for solving IFP in polynomial time, the RSA encryption algorithm is currently unbreakable in reasonable time. Since the AKS algorithm solves PRIMES and not IFP, it cannot be used to break RSA encryption (Cormen et al 2001, 881‐7).

2.5. Conclusion

To sum up this section, this is the state of affairs from the point of view of computational theorists. In the theory of computation, a distinction exists between problems that can be solved in reasonable time and those that cannot. In addition, a theoretical distinction exists between deterministic and probabilistic algorithms, but this distinction has no practical implications when algorithms are put to use. Two problems exist in the theory of computation with respect to prime numbers, PRIMES and IFP. The AKS algorithm, which was presented in the ‘PRIMES is in P’, is a deterministic algorithm that solves PRIMES, but not IFP, in reasonable time. There had already been at that time a probabilistic algorithm – the Miller‐Rabin algorithm – that solves PRIMES in reasonable time. As the Miller‐Rabin algorithm is reliably used for establishing RSA‐based encryption systems on the Internet, the introduction of the AKS algorithm does not change the way data are encrypted on the Internet. In order to break this encryption, an algorithm that solves IFP in reasonable time is needed. Currently, no such algorithm exists.

‐ 68 ‐

3. The General Press Coverage of ‘PRIMES is in P’

4 August 2002 Agrawal and his team email a draft of their paper to fifteen expert mathematicians and computer scientists. 7 August 2002 Agrawal and his team publish their paper on the Internet.

8 August 2002 The New York Times publishes an article about their paper.

9 August 2002 – Several other local and international newspapers pick up the story and publish articles about the paper. 26 August 2002 30 October 2002 Agrawal gives a talk at Clay Mathematics Institute in Cambridge, MA, and receives the Clay Research Award. 31 October 2002 – Agrawal gives talks at MIT, Harvard University and Princeton University.

11 November 2002 4 November 2002 The Wall Street Journal publishes an article about Agrawal’s talks.

November 2002 – The popular press publishes a few additional articles about the paper.

March 2003 24 January 2003 The paper is accepted for publication by the peer‐reviewed journal Annals of Mathematics. 27 May 2003 Agrawal receives the International Centre for Theoretical Physics (ICTP) prize. September 2002 Annals of Mathematics publishes the paper.

24 April 2006 Agrawal and his team win the Gödel Prize for their paper in Annals of Mathematics.

Table 1: 'PRIMES is in P' Timetable In an article in Notices of the American Mathematical Society, Folkmar Bornemann, a mathematician at the Technische Universität München, depicts the process of discovery and reception of the AKS algorithm and the basic mathematical ideas on which it is based. According to Bornemann, on Sunday, August 4, 2002, Agrawal and his two students sent a pre‐print of their article to fifteen experts by email. On Monday, Carl

Pomerance, a world‐renowned expert in number theory, confirmed the result, and informed the reporter Sara Robinson from The New York Times. On Tuesday, Agrawal and his students made their paper available online (Bornemann 2003, 545). On

‐ 69 ‐

Thursday, August 8, 2002, an article by Sara Robinson was published in the Science section of The New York Times under the title ‘New Method Said to Solve Key Problem in

Math’. The first paragraph of the article says:

Three Indian computer scientists have solved a longstanding mathematics problem by devising a way for a computer to tell quickly and definitively whether a number is prime — that is, whether it is evenly divisible only by itself and 1. (Robinson 2002, A20; emphasis added)

The words ‘quickly and definitely’ are prone to two different interpretations in two different contexts. Understood in their theoretical context, ‘quickly’ means

‘belonging to complexity class P’ and ‘definitively’ means ‘deterministic’. However, in ordinary language, ‘quickly’ is the opposite of ‘slowly’ and ‘definite’ is the opposite of

‘indefinite’. This gives to the lay reader the false impression that, until that time, there had been no ‘quick and definite’ algorithm for checking whether a number is prime

(recall, the Miller‐Rabin method does just that). Then the article states:

Prime numbers play a crucial role in cryptography, so devising fast ways to identify them is important. Current computer recipes, or algorithms, are fast, but have a small chance of giving either a wrong answer or no answer at all. The new algorithm — by Manindra Agrawal, Neeraj Kayal and Nitin Saxena of the Indian Institute of Technology in Kanpur — guarantees a correct and timely answer. (Robinson 2002, A20)

As before, the term ‘small chance’ has also very different meaning in the theoretical context and in the ordinary language context. As opposed to the actual state of affairs, the lay reader gets the impression that until now, there has been no reliable way to identify prime numbers. On the different interpretations of the terms in the article by mathematicians and the lay public, Bornemann remarks:

‐ 70 ‐

The … New York Times article celebrated the result as a triumph, but opaquely by choosing to simplify to a ridiculous extent: polynomial running time became ‘quickly’; deterministic became ‘definitively’. The article thus reads as follows: three Indians obtained a breakthrough because the computer could now say ‘quickly and definitively’ if a number is prime. On the other hand, the new algorithm has no immediate application, because the already existing methods are faster and do not err in practice. ‘Some breakthrough,’ readers would say to themselves. (Bornemann 2003, 550‐1)

Another way in which the New York Times article creates a false impression to the lay reader is by the juxtaposition of the issue of cryptography next to the sentence about the fact that previous algorithms have a ‘small chance’ for error. It is true that prime numbers play a ‘crucial role in cryptography’, but the article does not state what role. The word ‘so’ implies a causal relation between cryptography and the need to identify prime numbers, neglecting the fact that there already exists an efficient and reliable way – in the ordinary sense of these words – to achieve this task. By juxtaposing these two pieces of information, and by taking into account that the lay reader will misinterpret the meaning of ‘small chance’, the article creates a false impression to the lay reader. Without saying this specifically, the article gives the idea that until now there have been some problems with cryptography algorithms – perhaps they were not reliable enough because they sometime made mistakes – and that the new algorithm is going to fix this problem.

The NYT article was followed by several other reports in the general media. The

NYT article was first picked up by the Indian media. On August 9, the Indian daily

English newspaper The Hindu published the article, only changing the title to ‘New

Algorithm by Three Indians’, thus changing the emphasis to matters of national pride

(Anonymous 2002).38 On August 9, 2002 an article was published on the Internet sites

38 Other reports in the English language Indian press are Pradhan (2002), Rajghatta (2002) and

‐ 71 ‐

CNET and ZDNET with the title ‘Prime efforts may boost encryption’ (Junnarkar 2002).

In this article, the AKS algorithm is depicted as a possible improvement upon the popular and widely used RSA encryption algorithm, since old primality tests had a

‘miniscule probability of producing a wrong answer’ while the new algorithm ‘is believed to generate correct results each and every time’. In fact, in this article, Agrawal is quoted as saying: ‘Our algorithm is slower than the fastest‐known primality testing algorithms. The satisfying part of our algorithm is that it is completely deterministic as opposed to earlier ones that may make an error‐‐even though rarely’ (Junnarkar 2002).

In other words, what was only implied in the NYT article – namely, that the new

AKS algorithm is important because of its possible practical use in cryptography, became explicit. Agrawal’s own words, which have two very different interpretations in the theoretical versus common language context, seem to strengthen this conclusion, since the article does not mention that the fact that previous algorithms ‘make an error‐

‐even though rarely’ is just a theoretical characterization of these algorithms, which does not have any bearing on the practical reliability of these algorithms for encryption.

Later that year, an article in The New York Times Magazine reinforced this false interpretation of the AKS algorithm contributing to increased Internet security, and even to the war against terror(!):

Encryption programs used by banks and governments rely on increasingly large primes—up to 300 digits, these days—to keep criminals and terrorists at bay. This new algorithm could guarantee primes so massive they would afford almost perfect online security. (Thompson 2002, 107)

Ramachandran (2002). The latter is the only exception of an article from the popular press that I found during my research which gives an accurate and non‐distorted account of the importance of the AKS algorithm in its scientific context.

‐ 72 ‐

On August 14, 2002, an article was published on The Australian Broadcasting

Corporation Internet news site. This article also gave a wrong interpretation of the new algorithm, emphasizing its practical applicability for improving the RSA encryption algorithm, saying that existing algorithms for primality check were either inefficient or

‘carry a small degree of inaccuracy’. The article explained that the new algorithm was able to determine if a given number was prime, but not to factor a number. The article concluded with the commentary of Mr. Adam Spencer, ‘a mathematician by training’, saying: ‘If someone was to develop a program that was able to factor numbers, the whole security process of data would collapse’ (Kingsley 2002).

The ABC article managed to distinguish the PRIME problem from the IFP problem, but this was not the case for all the articles about the AKS algorithm. On

August 19, 2002, the Israeli daily newspaper Haaretz, a high‐quality broadsheet, published an article with the sensational title ‘The Prime Numbers Will be Identified, the Code will Be Broken’. The article, which references the NYT article from August 8,

2002, states that if the new algorithm developed by the three Indians really works, ‘it will be able to serve as an effective tool for breaking digital codes’. By confusing the IFP problem with the PRIMES problems, the article states that the current RSA encryption used on the Internet cannot be broken because ‘current methods for determining the primality of numbers are either too slow or not certain’, and concludes: ‘it will be shortly made clear if this is indeed a development which undermines the ability to encrypt digital data’ (Brizon 2002). About a week later, the paper published a correction article by Dr. Tamir Tassa, a mathematician from The Open University of

Israel, with the title ‘With all Due Respect to the Deterministic Algorithm in Polynomial

‐ 73 ‐

Time, the Code Will not Be Broken’. This article interpreted the AKS algorithm in its theoretical context (note also the use of mathematical jargon in the title!), explaining the reliability of the existing Miller‐Rabin probabilistic algorithm for checking primality and distinguishing PRIMES from IFP (Tassa 2002).

Between October 20 and November 3, 2002, Agrawal held a series of talks at

MIT, Harvard University, Clay Mathematics Institute in Boston and Princeton University.

Following this series of talks, an article was published in The Wall Street Journal. The article’s subtitle states ‘Will Manindra Agrawal bring about the end of the Internet as we know it?’. The article states that Agrawal found an algorithm for determining if a number is prime, and suggests that just another small step is needed to find an algorithm for factoring a large number – a development that would break Internet encryption. Describing the attention Agrawal got from computer scientists and mathematicians, the article states:

Prof. Agrawal's work involved only testing whether a number is prime, not the factoring problem. Still, there are enough connections and similarities between the two that mathematicians and computer scientists from all over the East Coast flew in to hear Prof. Agrawal on a whirlwind tour last week through the likes of M.I.T., Harvard and Princeton. (Gomes 2002, B1)

The article connects the wide academic attention Agrawal got to the fact that his algorithm might be used for breaking Internet encryption. However, a reading of the abstract of the paper Agrawal presented at MIT suggests that the academic interest in his work was purely theoretical:

‐ 74 ‐

Testing if a given number is prime is a fundamental and well‐studied problem of computational number theory. There are several algorithms known for it that are efficient in various ways: deterministic polynomial time under ERH (Miller), randomized polynomial time (Rabin, Solovay‐Strassen, Adleman‐Huang), and deterministic ‘slightly’ superpolynomial time. However, till recently, there was no unconditional deterministic polynomial time algorithm known for the problem. In this talk, we present the first such algorithm.39

Note that this abstract does not mention anything about Internet encryption or the IFP problem. It deals only with the PRIMES problem in relation to its classification to complexity classes. Of course, I am not denying that the overall interest in algorithms concerning prime numbers has to do with their use in encryption. However, as opposed to The Wall Street Journal’s allegation, ‘connections and similarities’ between PRIMES and IFP were not presented at Agrawal’s talks.

An article in The Economist from March 29, 2003 with the title ‘Primed to Go:

Mathematicians Are Discussing Ways to Make Code‐Breaking Easier’ is written along the same lines as The Wall Street Journal and states the following:

There is still some way to go before any of this work actually threatens cryptography. That is because quick and dirty techniques for testing primality already exist. Unlike Dr Agrawal's method, and its slower predecessors, these sometimes make mistakes, falsely attesting that a number is prime. But because such mistakes are rare, they are tolerable. However, if Dr Agrawal's primality test can be extended to factoring numbers, it would mean a rejigging of modern cryptography. Then the spooks and bankers really would be worried (Anonymous 2003, 89).

The word ‘because’ in the beginning of the second sentence creates an alleged connection of cause and effect between the fact the AKS algorithm does not yet pose a threat to internet security and the probabilistic nature of the existing Miller‐Rabin algorithm for testing primality. However, from the point of view of computational theorists, this is false. First, recall from section 2.4 that the existing Miller‐Rabin

39 .

‐ 75 ‐

algorithm is used for encryption and not for code breaking. Second, as stated in 2.2, the fact that it is probabilistic is a theoretical characterization of it, and has nothing to with its reliability for practical use.

The popular press, therefore, misinterpreted the AKS algorithm, its significance and implications. While mathematicians and computational theorists understood it as a theoretical breakthrough, the popular press emphasized its alleged practical significance for encryption on the Internet. The press misinterpreted the term

‘probabilistic’ when used to refer to algorithms as meaning ‘having a tangible probability of error in the practical domain of encryption’. By doing so it concluded that the former probabilistic Miller‐Rabin algorithm for checking primality was unreliable. It therefore interpreted the new AKS algorithm either as an algorithm that could improve current encryption on the Internet by making it more reliable, or as an algorithm that could be used for breaking the current method of encryption. Both of these contradictory interpretations are false from the point of view of specialists familiar with the problem.

A question then arises about what caused the popular press to misinterpret the significance of the ‘PRIMES is in P’ paper. In the next section, I will argue that this misinterpretation served both the interests of the popular press as well as the interest of the scientists who made the discovery. I will argue that the fact that the relevant scientific community had strict epistemic standards for evaluating the paper in question and distinguishing a correct from a distorted understanding of it, explains how Agrawal and his team were able to have different interpretations delivered simultaneously to

‐ 76 ‐

different audiences through different communication channels, achieving maximal exposure without risking their scientific reputation.

4. The Shared Interests of the Scientists and the Press in Distortion

Gregory and Miller identify several characteristic, or ‘news values’, which make stories appealing to media consumers. Stories about events that happen on a large scale have more news value than those that happen on a small scale. Stories relevant to readers’ lives, or stories about matters on which readers already have knowledge or opinions, have relatively high news value. Exclusive stories have more news value than stories that are widely accessible. Bad news is more newsworthy than good news. Readers have more interest in stories that happen in their own back yards than those that happen far away. Last, stories from reliable sources have more news value than stories from dubious sources.

As Gregory and Miller observe, with the exception of having what is perceived as a reliable source, scientific stories typically lack news value: they usually happen on a small scale; they touch on aspects that are foreign to people’s lives; stories about scientific discoveries are usually not exclusive; their immediate negative or positive impact is not clear; and they are often universal and not local (Gregory & Miller 1998,

110‐114). News reports about science therefore try to be as relevant and meaningful as possible: they make bold claims; they lack nuance; they emphasize the potential application and outcomes of scientific results; and they connect scientific results to matters that are close to the readers’ world (Gregory & Miller 1998, 116).

News values are a set of epistemic standards that are external to the scientific community. Scientists determine what knowledge of science the general public will

‐ 77 ‐

have only to a small extent. So do science journalists. News editors, who have no particular loyalties to science as an enterprise, largely determine which stories get published, and they typically publish what they believe the public wants to read

(Gregory & Miller 1998, 109).

If scientists have interest in the general media’s coverage of their work, as I will argue they do, they must cooperate with the media’s epistemic standards. At the same time, since their fellow scientists are also media consumers, they will not want the media reports to discredit their work in their colleagues’ eyes. But when distortion is easily distinguished from genuine scientific knowledge, they needn’t worry about this so much. They can enjoy both worlds.

These observations are important in analyzing the PRIMES affair. When asked by his fellow mathematicians about his impression of the popular media coverage of his discovery, Agrawal politely advised, ‘leave aside the general public coverage’

(Bornemann 2003, 550). However, although Agrawal was reluctant to comment about it to his colleague, he did cooperate with the popular media coverage.

Agrawal’s cooperated in a twofold manner. First, he gave interviews to reporters from the popular media. As I will show, it is likely that some of the journalists’ misconceptions originated from these interviews. Second, on the webpage where they published their original ‘PRIMES is in P’ paper, Agrawal and his students published several links to media reports about their paper, including the first NYT report.40

40 The original web page was in the URL , but it was no longer available online after April, 2005. However, it can still be accessed in the Internet Archive site in the following URL .

‐ 78 ‐

When interviewed by The Wall Street Journal, Agrawal made the following statement regarding his motivation to find a deterministic polynomial algorithm for solving IFP (recall that, if found, such an algorithm can be used to break the popular

Internet RSA encryption system):

‘Factoring is a natural problem. And natural problems should have a natural complexity to them. But this … is not natural complexity. This looks very strange. There must be something more natural than this out there’. (Agrawal, quoted in Gomes 2002, B1)

To the lay reader, such a statement may seem reasonable. However, to a specialist, it makes less sense, because the word ‘natural’ used in this context is vague and ambiguous. The term ‘natural problem’ does not refer to any particular class of problems. Does a ‘natural problem’ involve natural numbers? Perhaps, or perhaps not.

Likewise, the term ‘natural complexity’ is not associated with any of the known complexity classes. At most, if it is meaningful at all for a specialist, this statement may express some vague intuition and nothing more. However, to the lay reader, to the extent Agrawal is quoted correctly, this vague statement provides the missing speculative link between the actual AKS algorithm and ‘putting the Internet on alert’.

An examination of Agrawal and his students’ web page shows a conscious use of this medium. Their site contains three main sections, which target roughly three types of audience. Two sections, the first and third, have already been mentioned. The first section contains a link for downloading their original ‘PRIME is in P’ paper. This is obviously intended for the specialist audience of mathematicians and computer scientists. The third section is the list of links to popular reports about the algorithm.

‐ 79 ‐

The other section, the second on the page, contains a link to a list of Frequently Asked

Questions (FAQ) about ‘Prime is in P’ compiled by Anton Stiglic (2005).41

This FAQ targets an audience of people who are interested in theoretical computer science, but are not professional academics, such as students, software engineers, and amateur mathematicians. It aims at explaining the principles of the AKS algorithm and its importance. Specifically, it aims at clarifying common misconceptions about the algorithm, such as the confusion between PRIMES and IFP, as the following excerpt from it shows:

Q13. Does this result have any impact in cryptography at all? Not in any obvious ways. Certain algorithms need to generate prime numbers in order to construct cryptographic keys, but algorithms to accomplish this which can be executed very efficiently already existed before the result in [1]. The most commonly used ones have a probability of error, but this error can be made to be arbitrarily small … and thus they give us practically the same assurance as the algorithm proposed in P. These algorithms that are commonly used in practice are actually faster than the ones proposed in [1]. The result in [1] is a very important one in complexity theory, but probably have no (practical) impact in cryptography. (Stiglic 2005)

If we compare this statement to Agrawal’s statement in The Wall Street Journal we get quite a different impression. While according to this statement, AKS has no impact on cryptography, at least ‘not in any obvious way’,42 the impression we get from the article

41 According to his homepage, Stiglic is a cryptologist who obtained his M.Sc. in theoretical computer science from the Université de Montréal, and is also active at the ‘Crypto and Quantum info Lab’ at the School of Computer Science at McGill University . 42 When making claims, scientists tend to use cautious and modest language. This may be explained inter alia by their adherence to the Popperian ethos, in which all knowledge claims are provisional and may be subject to future falsification. Other epistemic communities, such as the popular media, prefer a much more decisive language. As Beecher‐Monas (2007) notes, in the legal system, this difference in tone has profound implications. While scientists tend to use cautious language, judges prefer the language of certainty. When encountering cautious language, judges often exclude scientific evidence as speculative. This shows their lack of understanding of the cultural norms of modesty and caution of the scientific community, and a failure to evaluate the evidence on its own stake (Beecher‐Monas 2007, 54‐55).

‐ 80 ‐

in The Wall Street Journal and the other articles in the general press is that just another small step is needed for AKS to be used to break Internet cryptography.

The conclusion of this comparison is clear. Different messages were simultaneously delivered to different types of audiences through different communication channels. It seems that it was very difficult for Agrawal to receive wide exposure in the popular media without implying that his algorithm has practical implications for Internet security. On the other hand, Agrawal could count on his colleagues to immediately recognize the real significance of the paper, which is purely theoretical, and simply dismiss the general media reports about the algorithm as popularized distortions. In other words, as it lacked news value on its own, it was very unlikely that the general media would report only a theoretical breakthrough in computer science if it had not been implied that this breakthrough has practical implications for

Internet security.43 On the other hand, these reports in the general media did not compromise Agrawal’s prestige among his peers because they could clearly distinguish genuine knowledge from distortion. This example shows that, as opposed to Bucchi’s claim (1996, 378), such simultaneous communication at different levels does not necessarily mean that barriers between genuine and popularized knowledge cannot be drawn sharply. Rather, it implies the contrary.

43 In addition to being consistent with Gregory and Miller’s analysis of news values, the plausibility of this claim is supported by historical research. Hughes (2007) describes the work of Manchester Guardian science journalist James G. Crowther in interwar Britain. While Crowther expressed interest in reporting about new developments in atomic physics and the discovery of new subatomic particles, his editor tended to perceive these issues as too complicated and lacking in interest for the newspaper’s readership. Instead, he encouraged Crowther to inform readers about mundane issue such as ‘eels, the physiological effects of manual labour, and dairy farming’ (Hughes 2007, 16). It was only Crowther’s success in achieving priority in reporting about the developments in atomic physics and competing journals’ consequent interest in these reports that persuaded his editor to approve their publication.

‐ 81 ‐

What did Agrawal and his team have to gain from the general media coverage of their algorithm? Or in other words: What were their interests in such coverage? What may explain their choice to publish their result on the Internet and to turn to the general media before pursuing regular channels of peer reviewed publications? An analysis of this case points to three interests: visibility, recognition and priority.44

4.1. Visibility

The popular media is an ideal means of getting as much exposure as possible. The first article in the NYT was a great promotion that attracted the attention of mathematicians and computer scientists. Within the first day that their paper was available online, it was downloaded by about 30,000 people (Kingsley 2002). Within the first ten days of being online, the dedicated webpage had more than two million hits and 300,000 downloads of the paper itself (Bornemann 2003, 546).45 The coverage of the paper in the general media was the trigger, or at least a major catalyst, for this extremely vast interest.

44 Fuller identifies a general interest of the scientific community in popularization. This is the interest in science’s continued survival. The scientific community has an interest in popularized accounts in the media, because they help science gain the support of the public, but at the same time they do not provide the public with sufficient in‐depth understanding of science to enable them to question scientists’ work (Fuller 1997, 32‐33). Within the scope of this chapter I will only discuss interests of individual scientists in popularization. 45 The ABC article states that the paper was published on the Internet on August 7, 2002, and ‘within 24 hours’ it was downloaded more than 30,000 times. The NYT article was published on August 8, 2002. So due to the time difference between Kanpur and New York, if we take the words ‘24 hours’ literally, and we start counting from the early morning of August 7 (Kanpur time), then these downloads occurred before the NYT article was published. However, if we start counting the hours from the evening of August 7 (Kanpur time), or do not take the words ‘24 hours’ literally, then it turns out – and this is the plausible scenario in my opinion – that the NYT article did contribute significantly to the number of downloads. Otherwise, this enormous number cannot be explained. The alternative explanation is that the rumour about the paper was spread by emails. However, if this was the case, members of the initial group of people that could spread this rumour had already had the article sent to them by email. It would have been more plausible that they would have forwarded the actual paper by email to their colleagues as well, saving them the need to download it from the Internet themselves. Therefore, it is much more plausible that the NYT article was responsible for the great number of downloads.

‐ 82 ‐

Another important reason for turning to the general media is to enhance the number of citations of an article. A study that compared the number of references in the

Science Citation Index of articles in The New England Journal of Medicine that were covered by The New York Times with the number of citations of similar articles that were not covered, shows that Articles in the Journal that were covered received a disproportionate number of scientific citations in each of the ten years after they appeared (Phillips et al 1991).46 It is reasonable to assume that the first article about the ‘PRIMES is in P’ paper in The New York Times was the most influential, because that newspaper is believed to set the tone for other general papers and magazines (Phillips et al 1991, 1180).

As of June 2008, Google Scholar counts about 126 references to the original

‘PRIMES is in P’ article that was published on the Internet in 2002,47 and 367 references48 to the peer‐reviewed article published in 2004 in Annals of Mathematics.49

The above study suggests that without the New York Times publication, it would have been lower.

4.2. Recognition

In this case recognition is a direct result of visibility, and the epistemic standards of evaluation of the relevant scientific community were rigid and well defined. Namely,

46 The researchers had a control group of articles published in a three month period in which The New York Times was on strike, which militates against the possibility that the articles that appeared in that newspaper were simply the most important ones. 47. The method I used was to count the number of references to the original URL in which the article was first published. 48 . 49 Unfortunately, the ISI Web of Science gives inaccurate results about this article. According to the Web of Science, the 2004 article was cited only 8 times! However, when I checked some of the citing articles that appeared on Google Scholar, I found that they did exist on the Web of Science, but for some reason did not appear among the articles citing the 2004 article.

‐ 83 ‐

there is a consensus among professional mathematicians and theoretical computer scientists about what constitutes a mathematical proof.50 Moreover, Agrawal and his students’ paper was short ‐‐ only eight pages long. Unlike other proofs, the math it used was relatively simple and accessible to an advanced undergraduate math student

(Bornemann 2003, 545).51 Thus, Agrawal and his team did not need to wait for the long and tedious peer review process. Thousands of professional mathematicians and computer scientists downloading the paper and checking the proof were better than

50 MacKenzie describes how the concept of proof in computer science and mathematics has changed in the second half of the twentieth century, side by side with developments in computer technology. Within mathematics and computer science, he identifies two main subcultures that have emerged. One subculture sees proof as a logical manipulation of symbols in a formal language that can, at least potentially, be performed by a computer. The second subculture sees proofs as rigorous arguments that can convince a trained human mathematician. While subscribers to the former view will tend to regard proofs that appear in textbooks and academic papers, such as the proof that PRIMES is in P, as sketches of formal proofs, subscribers to the latter view will tend to regard formal proof as a partly adequate and idealized model of real, rigorous argument proofs (MacKenzie 2001, 323‐24). MacKenzie argues, however, that these two views are not incompatible enough for actual mathematical proofs to genuinely constitute what Galison (1997, Ch. 9) calls a ‘trading zone’, namely a site where diverse cultures coordinate their practical activities while maintaining a distinct understanding of the meaning of what they do and what they exchange. Different types of proof that conform to different perceptions of what a proof is are allowed to live peacefully together in the mathematical literature and are rarely disputed (MacKenzie 2001, 327‐8). As MacKenzie points out, while there no one agreed upon view among mathematicians about what exactly a mathematical proof is, ‘this does not imply that “anything goes,” that any arbitrary argument can count as a mathematical proof. What it suggests, rather, is that members of the relevant specialist mathematical community, in interaction with one another, come to a collective agreement as to what counts as a mathematical proof’ (MacKenzie 2001, 318). Moreover, MacKenzie’s research did not find a case in which a mechanical proof disagreed about a theorem with an established rigorous proof that had preceded it (2001, 323). We should also distinguish between disagreements on the nature of proof (epistemic standards) from disagreement on the truth and falsity of theorems (knowledge claims). For example, in the case of the four colour theorem, which was controversially proven with the aid of a computer program, if we ignore the groundless rumours about a bug in the program that was used to prove it (MacKenzie 2001, 139) mathematicians do not dispute whether the theorem itself is true, only whether the method that was used to show its truth constitutes a proof. 51 Agrawal and his students’ proof was an ordinary mathematical proof like the vast majority of mathematical proofs that appear in mathematical journals and textbooks. It did not involve the use of computers. Thus, unlike the proof of the four colour theorem, for example, which relied extensively on the use of computers, the proof that PRIMES is in P did not trigger debates about its validity. In addition, because Agrawal and his students’ proof was relatively short and non‐complex, it did not trigger debates such as in the case of the proof of Fermat’s last theorem, the length and complexity of which made it difficult to be verified.

‐ 84 ‐

any peer review process. Of course, not all cases are like that, and visibility does not always result in recognition.

Another important aspect of recognition is prizes. An article on the Internet news site rediff.com reports that ‘IIT Kanpur Director Sanjay Dhande was elated at the news that created headlines in The New York Times’. A few sentences later it adds: ‘He was confident about Agrawal getting nominated for the world's top honours in mathematics, considering his latest feat’ (Pradhan 2002). The connection between media coverage and recognition in the form of awards was not concealed to IIT Kanpur director Dhande, and indeed on October 30, 2002, less than three months after the initial publication of Agrawal’s paper on the Internet, and before his paper was published in any peer‐reviewed journal, Agrawal won the Clay Research Award at the

Clay Mathematics Institute in Cambridge, Massachusetts. In May 2003 he won the

International Centre for Theoretical Physics (ITTC) Prize. In April 2006, after the article appeared in Annals of Mathematics in September 2004, Agrawal won the prestigious

Gödel Prize, which is given only to articles published in peer‐reviewed publications.

4.3. Priority

Computer science is a field with very rapid developments. It may be the case that by the time an article is published in a peer reviewed journal, it is already outdated. The publication of ‘PRIMES is in P’ in a peer reviewed journal (Annals of Mathematics)

(Agrawal, Kayal & Saxena 2004) occurred more than two years after the paper appeared online. It was almost nine months from the moment the paper was accepted to the moment it was published. The slow process of peer‐review is incompatible with

‐ 85 ‐

the fast pace of the field. As Odlyzko, writing before the publication of the paper in

Annals of Mathematics, observes:

The [peer‐reviewed] journal version will probably be the main one cited in the future, but will likely have little influence on the development of the subject. Within weeks of the distribution of the Agrawal‐Kayal‐Saxena article, improvements on their results had been obtained by other researchers, and future work will be based mainly on those. Agrawal, Kayal, and Saxena will get proper credit for their breakthrough. However, although their paper will go through the conventional journal peer review and publication system, that will be almost irrelevant for the intellectual development of their area. (Odlyzko 2003, 311).

Because of the dynamic nature of this field, it is plausible to assume that scientists in it will give prime importance to the issue of priority. Publishing a paper on the Internet is the best way to win a priority race. Because there is a consensus among scientists about what constitutes a mathematical proof and presenting such a proof was all that was required to achieve acceptance of their claim, Agrawal and his team had nothing to lose by their Internet publication and their use of the general media, only to be acknowledged as the first to find a deterministic polynomial algorithm for PRIMES.

5. Popularization and Distortion Revisited

A debate exists about the question of media popularization. The question is whether there is ‘genuine knowledge’ in contrast to ‘distorted knowledge’. In this chapter I have shown that in the PRIMES affair, the general media did give a distorted account of the

AKS algorithm, its importance and implications. These modes of distortion, several of which were usually used together, are summarized in Table 2.

‐ 86 ‐

Mode of Distortion Examples 1 Using terms which have different meaning ‘The new algorithm — by Manindra Agrawal, in ordinary language context and in Neeraj Kayal and Nitin Saxena of the Indian scientific context; Institute of Technology in Kanpur — guarantees a correct and timely answer.’ (Robinson 2002, A20; emphasis added) 2 Neglecting to mention facts that are Many reports ignored the fact that a reliable relevant for understanding the algorithm for testing primality is already used in significance of the discovery; RSA encryption. 3 Obscuring the difference between similar Obscuring the difference between PRIMES and IFP. but not identical things; 4 Mentioning two facts which are true in ‘RSA, a popular encryption algorithm used in themselves, but juxtaposing them or securing Internet commerce, is built on the making a logical connection between them assumption that when prime numbers … are large in a way that creates a false impression enough, they're nearly impossible to generate and that they are connected; determine. … But a new algorithm, developed at the Indian Institute of Technology in Kanpur by Manindra Agrawal and his students Neeraj Kayal and Nitin Saxena, is believed to generate correct results each and every time.’ (Junnarkar 2002) 5 Using speculative language, e.g., words like ‘Prime Efforts May Boost Encryption’ (Junnarkar ‘possibly’ or ‘may’, phrasing sentences in 2002); the form of a question, etc., while the ‘Will Manindra Agrawal bring about the end of the speculations are unfounded; Internet as we know it?’ (Gomes 2002, B1). 6 Making false statements. ‘It will be shortly made clear if this is indeed a development which undermines the ability to encrypt digital data’. (Brizon 2002) Table 2: Modes of Distortion by the Media of the ‘PRIMES is in P’ Paper In the usual case, a reader who is familiar with the relevant scientific discourse could find a kernel of truth in a media report, or at least identify the genuine fact on which the distorted account is based. However, in rare cases, there is no interpretation, not even one extremely liberal and charitable, under which statements in the media report may be considered to be even partly true. This is the case that corresponds to mode of distortion (6). The example which is quoted in (6) is the only case I have found during my research in which a newspaper published a correction.

Viewed from the perspective of computation and number theorists, the distorted accounts in this case could be distinguished from faithful representations on stable and recognizable epistemic grounds. The existence of these stable communal epistemic

‐ 87 ‐

grounds explains Agrawal and his students’ ability to deliver different accounts, distorted and non‐distorted, to different audiences, thus achieving both visibility and professional recognition at the same time. In this case, the standards of the scientific community clearly and rigidly defined the borderline between genuine scientific knowledge and distorted simplifications.

I have argued that the existence of independent epistemic standards for evaluating knowledge claims explains the turn of events in my case study. One may argue, however, that my case study supports only a weaker claim than the one I have made, namely that researchers in a particular community are able to distinguish knowledge claims that adhere to their community consensus, where a consensus exists, from knowledge claims that depart from it. Hence, so this objection goes, what I call the independent epistemic standards merely turn out to be the standards of the particular community.

My response to this worry is threefold. First, even if we grant only my weaker claim, it still calls for a significant reform to the new model of popularization. Recall from section 1 of this chapter that according to the new model, scientists’ ability to label different accounts as legitimate or distorted science in different social circumstances is a political resource available to them to maintain their hegemony on the construction of scientific facts. This assumed ability is used to explain how certain claims gain the social status of knowledge. In order to effectively use this ability, the community’s epistemic standards need to be flexible and easily redefinable. The new sociological model of popularization suggests that when encountering a scientific report in the general media, individual scientists can make ad hoc changes to their epistemic standards either to

‐ 88 ‐

legitimize or discredit the report according to their social interests. If, however, epistemic standards are rigid and constrained by the pre‐existing community consensus, scientists cannot legitimize or discredit reports as they wish. It follows that scientists cannot maintain their hegemony on the construction of scientific facts as effectively as the new model suggests. This seems to seriously impair the potential explanatory potential of the new model.

Second, a stable communal consensus, where one exists, is itself an intriguing social fact that requires an explanation. How and why does such a consensus emerge and why is it maintained? A desideratum for a robust theory of popularization is to be able to explain the different outcomes, such as a change or lack of change in a community’s consensus, in similar cases in which scientists turn to the media. In particular, when does a scientist’s mere choice to turn to the media threaten the community to which she belongs and cause its members, for example, to penalize the deviating scientist by delegitimizing the factual claims in question? What distinguishes the cold fusion affair, in which media reports allegedly constructed scientists’ epistemic expectations (Lewenstein 1995; Simon 2001),52 from the PRIMES affair, where this was not the case? The new model does not seem to give us answers to these questions, whereas the introduction of the existence of independent and recognizable epistemic standards may explain the stability of a consensus in some cases.

Specifically, adherence to such standards may adequately explain why certain popularization episodes have so little influence on the shared beliefs of experts, while other episodes seem to have lasting effects on the course of scientific research. Of

52 For an alternative interpretation of the cold fusion affair, see Solomon (2001) at 129‐32.

‐ 89 ‐

course, it is not sufficient simply to assume that such independent epistemic standards exist. If the stability and instability of communal consensus is to be explained, subsequent sociological accounts of scientific knowledge will need to tell us when it is legitimate to make such an assumption and what it includes. The modest goal of the current study is to underscore the need for this kind of shift in explanatory strategy, rather than to propose a full‐blown methodological alternative.53

Third, the claim that the epistemic standards reflect merely local consensus overlooks an important aspect of my case study. The subject matter of Agrawal and his students’ claim was not restricted to the inner discourse of an esoteric group of number theorists. Rather, it was also a claim about what computers, material artefacts in the world, are capable and not capable of doing and at what speed. According to many of the newspapers reports, the AKS algorithm could be used to break the commonly used

RSA encryption on the Internet. Specialists, on the other hand, regarded this claim as false. To date, there have been no reported cases in which the AKS algorithm was successfully used to break RSA encryption in order to steal, for example, credit card numbers that are used in online transactions.

To date, then, in spite of opposing predictions in the general media, Web users can still safely use their credit cards and access their bank accounts online. What explains this ability? Why hasn’t the Internet collapsed, as predicted by some media reports? How are the inner social norms and conventions of an esoteric group of number theorists translated to the social norms and conventions governing online commerce? If we assume that the epistemic standards of the esoteric community of

53 See Tucker (2003) for such a discussion of scientific consensus.

‐ 90 ‐

number theorists are more than merely social norms, and that they are independent of the social settings in some way – that they are constrained by some physical and logical necessities, for example – the fact that the general media’s apocalyptic predictions about the collapse of the Internet have been proven false is easily explained. If we do not assume that, then we need a much more complex explanation which refers only to social norms. The current sociological model of popularization does not provide us with such an explanation, and I doubt very much if it can.

My claim is that the possibility of the existence of independent epistemic standards needs to be added to the explanatory toolbox of the new model of popularization. Cases like PRIMES cannot be adequately explained otherwise. Indeed, the absence of this assumption may reveal an explanatory lacuna in other case studies as well. For example, Sommer (2006) examines the discovery of a Neanderthal skeleton by French archaeologist Marcellin Boule in 1908. Boule interpreted the Neanderthal skeleton as a ‘cousin’ of modern humans, namely as having a common ancestry but not as a direct ancestor. Sommer, who relies on the new model (2006, 216) suggests that

Boule chose this interpretation because it was susceptible to two opposing popularized interpretations. The first interpretation suited the worldview of the secular progressivist French press, and the second was ‘Church friendly’ (2006, 231). Sommer does not examine, however, the extent to which the anatomical and paleontological beliefs of scientists of the time constrained the possible interpretations of the skeleton to begin with. If my argument in this chapter is correct, a complete explanation of the case should have addressed this as well.

‐ 91 ‐

While the possibility of the existence of independent epistemic standards is a nice thing to have in the explanatory toolbox of the new model, I do not argue that it should always be used. I do not claim that independent epistemic standards always exist, nor that when they exist, they always explain the outcome of a scientific affair.

Moreover, I do not claim that when independent epistemic standards exist, they necessarily correspond to the standards of the relevant scientific community. I call for a reform to the current model of popularization, not a return to old models of explanation, in which scientists’ true beliefs are explained by their truth, and false beliefs are explained by ‘external’ social factors.54

Developing such a theory of popularization exceeds the scope of this chapter, the main aim of which is to point out the need for such a theory. Nevertheless, I would like to make a few preliminary remarks about the generalizability of this case study. The case study I presented in this chapter was from theoretical computer science. However, the social mechanism I have identified and the distinction between genuine and distorted scientific knowledge is relevant to other cases as well. For example, a discovery of a new large asteroid is not news. If the asteroid is about to hit Earth, however, this is news. Therefore, in order to reach the general media, a scientist has every interest in overestimating as much as possible the chances that this newly

54 An example of the direction I am proposing is Solomon’s Social Empiricism. Solomon argues that social empiricism can offer symmetrical explanations for true and false beliefs, which invoke empirical, social, theoretical and cognitive factors (2001, 117‐20). For Solomon, empirical success is the main epistemic criterion for evaluating scientific theories (Solomon 2001, Ch. 2). The details of Solomon’s account may, of course, be debated. One may wonder, for example, whether empirical success is or should be the main epistemic criterion for theory evaluation in all social contexts. Additionally, it is not clear that empirical success is as independent of the social context as Solomon maintains. Nevertheless, the general principles of her overall framework may provide a promising avenue for developing a richer and more robust theory of popularization that accommodates the existence of independent epistemic standards and the ability to distinguish distorted from non‐ distorted scientific accounts.

‐ 92 ‐

discovered asteroid will hit Earth. Similarly, in the cold fusion affair, it is doubtful if

Fleischmann and Pons would have reached the general press had they not claimed the discovery of an extremely cheap energy source.55

At first glance, the PRIMES case study may seem different from perhaps more typical cases of popularization in that it involves popularization of mathematical knowledge, for which there are arguably rigid and stable epistemological standards.

This suggestion is problematic for two reasons. First, a principled distinction between mathematical and other forms of scientific knowledge conflicts with the epistemological commitments of the proponents of the new model. From its early days, SSK theory has denied any such principled distinction, arguing that the same kind of sociological analysis applies both to the ‘hard’ and ‘soft’ sciences (Collins 1983, 278).56 Proponents of the new model of popularization embrace this view, and regard the new model as part of the research programme that was set forward by the early SSK scholarship

(Whitley 1985, 6; Hilgartner 1990, 522‐4; Lewenstein 1995, 407). Therefore, they cannot simply explain away PRIMES as an unrepresentative exception to their model.

Second, as I have mentioned, the claims in the PRIMES affair are also about what computers, which are material artefacts in the world, are capable of doing.57 In this

55 As Pinch (1985) points out, scientists have a choice about how to put their claims. The more dramatic and less cautious they put them, the more they can gain in terms of reputation and recognition if their claims are ultimately accepted, and the more they can loose if their claims are ultimately not accepted. 56 For example, according to Bloor’s influential Strong Programme, the status of logical necessity or a priori knowledge is given to knowledge claims (at least primarily) thr ough social negotiations. Mathematical knowledge that has gained a secure status in the p ast can be occasionally challenged, and whether it retains its secure status is subject to a collective decision of the relevant epistemic community (Bloor 1991, 84‐156). Moreover, according to Bloor (1984), so‐called objective epistemic standards, including mathematical rules of inference, are the intersubjective social ly given meanings and categories in a given epistemic community, and are relative to it. Consequently, sociologists should analyze them in terms of the community’s social structure and collective social interests. 57 More precisely, the claims are about what a Turing Machine, which is an abstract model of a

‐ 93 ‐

sense, they are similar, for example, to physicists’ claims about what certain objects in the world can do, which are also expressed in mathematical language. What distinguishes this case from others, in my opinion, is that in PRIMES, the relevant scientific community required no empirical demonstration to secure the knowledge claim. Agrawal and his team were not required, for example, to code a computer program which solves PRIMES and test its performance. Presenting a proof was enough.

In particular, since ‘PRIMES is in P’ was a short and non‐complex paper, the correctness of the result could be easily verified. In computational theory, empirical demonstration in the form of coding a program and running it is considered irrelevant for supporting claims. This is not the case in other branches of computer science. For example, in computational linguistics, it is not enough to develop a new algorithm for speech recognition; it must also be empirically shown that it correctly recognizes speech.

Computational computer scientists’ confidence in the correctness of their claims without need to empirically demonstrate them may have many explanations. But, the enormous past success in easily implementing theoretical results such as the RSA encryption algorithm and the Miller‐Rabin primality testing algorithm surely has something to do with it.

In the cold fusion affair, in contrast, repeatable empirical demonstrations were required to establish the claim. Scientists who were trying to replicate Fleischmann and

Pons’ experiment and produce cold fusion learned many details about the experimental design from the media (Lewenstein 1995; Simon 2001), and so media reports obviously played a major role in mediating the knowledge claims at stake. However, note that this computer, can do. A Turing machine is considered equivalent in its asymptotic computational power to a digital computer because a digital computer can simulate a Turing machine in polynomial time and vice versa (Hopcroft et al 2001, 355‐65).

‐ 94 ‐

analysis of the role of the media shifts the focus to explaining why certain experiments failed or succeeded. This brings us back to notions of accuracy in reporting and distortion, which are associated with the old model, and not so much to social factors such as the prestige and reputation of the informants.

If this is the case, then, it seems that when replications are required and performed, such as in the cold fusion affair, or when experiments are irrelevant for establishing claims, such as in PRIMES, it is likely that the media will play a minor role, if any, in the construction of scientific knowledge. It seems more plausible that the media will play a more crucial role in constructing scientific knowledge in cases where scientists report certain empirical results, but attempts to replicate their experiment or reproduce their results are not performed due to various reasons such as cost and lack of resources. These are of course tentative suggestions that call for more sociological research.

Conclusion

In this chapter I have systematically surveyed and analyzed the popular press coverage of the ‘PRIMES is in P’ affair. I have argued—against the prevailing SSK orthodoxy—that without assuming that distorted simplifications of scientific knowledge are distinguishable from non‐distorted simplifications on independent epistemic grounds, the turn of events in the PRIMES case study cannot be adequately explained. This suggests that the possibility of the existence of independent epistemic standards must be acknowledged by SSK proponents.

‐ 95 ‐

Chapter 3

Knowledge­Based Consensus

Introduction

One of the main lessons we may draw from the previous chapter is that knowledge and social agreement in a community are not the same. As I have argued, the SSK view of knowledge as mere social agreement in a group faces difficulties with explaining patterns of stability and change in the beliefs of the members of the group. As the previous chapter demonstrates, the best explanation of the change in the mathematicians’ and complexity theorists’ beliefs following the announcement of the

PRIMES discovery was that they could appeal to epistemic standards that were independent of their social interests and goals.

Another lesson we may draw from the previous chapter is that scientists are not always reliable informants. As I have shown, the scientists in PRIMES conveyed a distorted account of the meaning and implications of their discovery to the general press in order to promote their social interests. We may therefore wonder under what general social conditions we should and should not trust scientists, other specialists, and the scientific community as a whole. This is, of course, a big question that requires more space than this dissertation allows. But we can narrow the scope of this question and identify particular cases in which an epistemic community is likely to be trustworthy.

As I have stressed in the previous chapter, although the consensus that emerged in the mathematical community after the announcement of the PRIMES discovery was based on shared knowledge, it would be generally wrong to assume that any consensus

‐ 96 ‐

in a community of specialists is based on shared knowledge. It is thus worthwhile asking under what conditions we would expect a consensus to be based on shared knowledge of it members. This is the question I address in this chapter.

The existence of a consensus in a community of researchers has long been perceived as a mark of knowledge. For example, Roman physician and anatomist Galen writes that there should be no quibbles between philosophers, in the same manner mathematicians are able to reach a consensus – the mark of true knowledge (Galen

1963, 58).58 Immanuel Kant writes that unlike logic, the field of metaphysics has not yet matured to be a proper science because metaphysicians are unable to reach consensus

( 1787/2007 , Bvii‐xxxv). In our times, bodies such as the National Institute of Health

(NIH), through its Consensus Development Program,59 and the Intergovernmental Panel on Climate Change (IPCC)60 are in the business of formulating expert consensus statements, which purport to provide authoritative answers to a variety of disputed questions. The existence of consensus in a scientific community is also used as a resource for arbitrating between rival experts in legal trials and making public policy decisions.

At the same time, a long tradition in philosophy is suspicious of consensus, and advocates the epistemic significance of dissent. By this tradition, too much agreement may be a mark of stagnation. For example, Mill argues that the existence of rival views to ours is necessary for correcting our beliefs when they are wrong and being able to justify them when they are right (1859/1993, 83‐123). Feyerabend argues that a

58 I thank Jackie Feke and John Christopoulos for these references. 59 http://consensus.nih.gov/ABOUTCDP.htm 60 http://www.ipcc.ch/ipccreports/ar4‐syr.htm

‐ 97 ‐

constant stream of new rival theories is essential for the advancement of science and human flourishing (1970, 209‐214), and Rescher (1993) argues against associating consensus with truth or taking it to be the aim or end of inquiry.

Recently, Feldman has argued that in light of the ubiquity of robust disagreement on many cardinal issues between informed and intelligent people, we know less than we think we know. He calls this ‘contingent real‐world scepticism’

(2006, 217). Similarly, Brian Frances coins the term ‘live scepticism’, to argue that in light of disagreement between experts on many issues, we ought to doubt many of our ordinary beliefs which we count as knowledge (2005, viii‐ix). By itself, however, the mere existence of agreement, either between experts or intelligent and informed people, is not necessarily a mark of shared knowledge. The existence of agreement as opposed to disagreement is a contingent fact, and people may come to agree for all kinds of reasons. This means that the same kinds of sceptical doubts that arise in the face of disagreement may arise also at the face of agreement. So do we know even less than we think we know?

This chapter addresses these puzzles. I address the question of when we may legitimately attribute a consensus in an epistemic community to knowledge that is shared by its members. I call such a consensus a ‘knowledge‐based consensus’. More precisely, I identify sufficient conditions under which it is very likely that a particular consensus in a real epistemic community is knowledge based.

I will make a twofold argument: (a) a consensus is knowledge based when knowledge is the best explanation of that consensus; (b) knowledge is the best explanation of a consensus when three conditions obtain:

‐ 98 ‐

1. The Apparent Consilience of Evidence Condition – the consensus exhibits an

apparent consilience of different lines of evidence, i.e. it seems epistemically

robust;

2. The Social Diversity Condition – the consensus is socially diverse;

3. The Meta­Agreement Condition – all parties to the consensus are committed

to the same general evidential standards, i.e. they agree as to what they

agree.

As I will explain, I take (a) to be the relatively unproblematic part of my argument, and I will spend most of this chapter defending (b) and the three conditions specified in it.

For the sake of clarity, let me stress what I do not argue in this chapter. I do not argue that consensus should be the aim of science. I do not specify general norms or procedures scientists should follow in their research. I do not even specify norms they should follow in case they want to reach a consensus. My aim is different. Often in public debates, such as in the debate on global warming or the Bendectin controversy, which I discuss in the next chapter, deference to consensus is done in order to resolve a certain dispute. My motivation for this chapter is to examine under what conditions such a practice is legitimate, hence I address the question of when a consensus is knowledge based.

This chapter consists of six sections. In section 1, I explore the relations between knowledge, luck and inference to the best explanation, and argue that a consensus is knowledge based when knowledge is best explanation of it. In section 2, I discuss my methodological assumptions, and distinguish my discussion of the epistemology of

‐ 99 ‐

consensus in real‐life cases from the discussion of the epistemology of consensus in idealized communities. In sections 3, 4, and 5 respectively I discuss my three conditions for knowledge‐based consensus. In section 6, I contrast my account with Longino’s

Critical Contextual Empiricism, and argue that Longino’s account of consensus suffers from several shortcomings that my theory avoids.

1. Knowledge, Luck and Inference to the Best Explanation

Some sociologists of knowledge and social epistemologists, such as Martin Kusch (2003,

62‐75), define knowledge as whatever an epistemic community comes eventually to agree on. If knowledge is consensus by definition, the question of when a consensus is knowledge based does not arise. As I have argued in Chapter 2, there are compelling reasons to reject such theories that conflate knowledge and consensus. For the purpose of this chapter, then, I will assume that knowledge is separable from consensus.

For many years, it was mostly uncontroversial that knowledge is justified true belief. This tripartite definition was challenged most forcefully by Gettier’s famous

(1963) paper, in which he argued that one can justifiably believe a falsehood from which one deduces a truth, thus having justified true belief but not knowledge. Gettier’s paper sparked a prolific period in which new theories of knowledge that were supposed to overcome the Gettier problem were proposed. These theories try to add a fourth condition or dispense with the notion of justification altogether (Blaauw & Pritchard

2005, 66‐9).

Broadly speaking, post‐Gettier theories of knowledge are dividable into two camps. Internalist theories analyze justification in terms of an agent’s mental state and her having good grounds for holding a defensible belief, while externalist theories

‐ 100 ‐

define external conditions that beliefs or belief‐forming processes need to meet. Notable among externalist theories are different versions of reliabilism61 and modal theories that define knowledge in terms of counterfactual conditions that beliefs need to meet to count as knowledge.62 Modal and reliabilist theories are typically compatible, as reliable processes of belief formation tend to meet the specified counterfactual conditions

(Blaauw & Pritchard 2005, 137).

For the purpose of analyzing the question of when a consensus is knowledge based, I need not adhere to any one theory of knowledge in particular (cf. Tucker 2003,

502). As I have argued in Chapter 1, what is significant for questions of evaluating the epistemic justificatory status of our real‐world beliefs is not which theory of knowledge or justification – internalism or externalism – is correct, but rather that whatever theory we employ will regard knowledge as a communal rather than individualistic good.

Moreover, what is significant for my account is not the many questions on which these theories diverge, but what they have in common.63 A central problem that Gettier‐ type problems expose, with which both internalist and externalist theories of knowledge and justification try to deal, is the problem of lucky beliefs. For example, if I form a belief by looking at a broken watch that happens to show the right time, although my belief is true and arguably justified, it is not knowledge. It is just a fluke. As Dancy

(1985, 134) writes: ‘justification and knowledge must somehow not depend on

61 See Goldman (2009) for an overview. 62 See, for example, Nozick (1981), DeRose (1995), and Sosa (2007). 63 As Foley (2005, 314) notes, while externalist and internalist theories of knowledge may be seen as rivals, they may also be seen as complementary, having different explanatory interests and trying to cash out the same notion of justification in different terms. Similarly, Feldman and Conee (2004, 94‐ 6) do not rule out that externalist and internalist theories may be extensionally equivalent. The view that an adequate theory of knowledge should incorporate externalist and internalist components is increasing in popularity, though how this should be done, of course, is much disputed (Prichard 2005, 67‐122; Goldman forthcoming; Comesana forthcoming).

‐ 101 ‐

coincidence or luck. This was just the point of the Gettier counter‐examples’.64 Pritchard calls the type of luck involved in Gettier‐type cases veritic epistemic luck, which is that ‘it is a matter of luck that the agent’s belief is true’. In modal terms, a veritically lucky belief is true in the actual world but false in a wide class of nearby possible worlds in which the agent forms her belief in the same way (Pritchard 2005, 146). For example, my belief about the current time will be false in most possible worlds in which I base it by looking at the broken watch. For a consensus to be knowledge‐based, then, it cannot be vertically lucky.

There is another type of epistemic luck, the elimination of which is central to any account of knowledge. I call it epistemic misfortune. It refers to circumstances in which an agent or an epistemic community are justified or seemingly justified in their beliefs but there are factors that systematically and/or deliberately mislead them or inhibit their gaining knowledge. Classical sceptical scenarios such as being deceived by a

Cartesian demon, being a brain in a vat or living in the Matrix world are examples of epistemic misfortune, as in all of these cases despite their best efforts, agents are systematically and/or deliberately being prevented from gaining knowledge of the world.65

Epistemic misfortune is not limited to far‐fetched sceptical scenarios, though.

There are plenty of cases of epistemic misfortune in the actual world. Such are cases in

64 See Pritchard (2005) at 125‐133 for a discussion of the conceptual relations between luck, chance, coincidence and accident 65 One may wonder how my own account deals with such sceptical scenarios. For example, suppose that there is a consensus in an epistemic community that Barack Obama is the president of the US, but in fact, he is not, and all the members to the consensus have been deceived by a Cartesian demon to believe so. Recall that my account is based on inference to the best explanation, which some philosophers take to be an adequate response to this form of scepticism (e.g. Vogel 1990). To the extent one thinks the inference to the best explanation response to scepticism cuts ice, my account successfully deals with such scenarios. In any case, my primary focus is ordinary world scenarios, in which such far‐fetched sceptical worries do not arise.

‐ 102 ‐

which the available evidence points at the wrong direction, when agents share a bias that interferes with their evaluation of the evidence, when different biases push different agents to the same wrong belief, etc. In some cases, careful reflection may lead agents to realize that they are in circumstances of epistemic misfortune,66 for example that their beliefs are biased by sexist or racial prejudice. In such cases agents may revise their beliefs accordingly.67 Typically, though, if the epistemic misfortune is genuine, they will not be motivated or disposed to carry out such reflection, and even when they will, they may fail. For a consensus to be knowledge‐based, then, it cannot be a victim of epistemic misfortune.

Epistemologists, interested in conceptual analysis of knowledge, have not directed much attention to epistemic misfortune, and unlike veritic luck, they have not tried to rule it out from their conceptions of knowledge. This is probably because when agents are epistemically misfortunate, their beliefs are false. Hence, any conception of knowledge as true belief will have already ruled out their beliefs as knowledge. When agents are vertically lucky, on the other hand, their beliefs are accidentally true, and

66 Pritchard argues that the type of epistemic luck involved in academic sceptical scenarios is reflective epistemic luck, which is that ‘given what the agent is able to know by reflection alone, it is a matter of luck that her beliefs are true’ (2005, 175). In academic sceptical scenarios, the evidence available to agent by reflection underdetermine her ordinary beliefs about the external world, as they can also support a scenario in which they are false and she is a brain in vat, etc. Thus, her beliefs are a matter of reflective epistemic luck (Pritchard 2005, 205). Unlike my notion of epistemic misfortune, however, reflective epistemic luck is applicable only to cases of academic scepticism, and not to cases of ‘contingent real‐world scepticism’ (Feldman 2006, 217), which are the focus of this chapter. In such cases, the source of the scepticism is not an agent’s inability to rule out far‐fetched scenarios based on her evidence, but sceptical worries about the contingent circumstances that have led her to her belief, which, as I have mentioned, are not necessarily unavailable to her upon careful reflection. 67 This sort of reasoning leads Fricker (2007) to argue for the centrality of the moral‐epistemic virtue of reflexive critical openness in epistemology in general and social epistemology in particular. I have addressed her argument more in depth in Chapter 1.

‐ 103 ‐

need to be ruled out for being a fluke.68 As opposed to epistemologists, philosophers of science, who are interested in the different ways in which beliefs or theories may seem or be justified or rational and still be false, have concerned themselves with cases involving epistemic misfortune, though have not explicitly invoked this term.

For the question of the conditions under which a consensus is knowledge based, the question of the conditions under which the parties to the consensus are epistemically misfortunate should obviously be addressed. If we only say that a consensus is knowledge based when it is not epistemically misfortunate, this amounts to trivially saying that it is knowledge based when it is justified and true.

In cases of epistemic misfortune, the parties to the consensus are victims of the circumstances. Despite their best efforts they are prevented from reaching warranted beliefs. By contrast, there are cases in which the parties to the consensus are less innocent – they deliberately form a consensus although they are aware, at least to some extent, that the view on which they form the agreement is not warranted and falls short of knowledge. These are cases of deliberate non­epistemic consensus, and are not knowledge‐based.

There are many reasons for people to deliberately form a non‐epistemic consensus. Consensus or the appearance of it is a powerful political tool for advancing policy and promoting social aims. A group of experts may have many reasons to reach a consensus and mask existing disagreements within it even when its members are aware of their existence. The experts may paternalistically decide it is better for the public that they speak in one voice; they may be afraid that their social status might be undermined

68 I thank Hagit Banbaji for this observation.

‐ 104 ‐

if disagreements among them became public; they may wish to gain external moral and material support; etc. In such cases, dissenters may choose not make their views known outside the community of specialists to which they belong in order to promote its shared collective interests (Beatty 2006, 53‐6; Fuller 2007, 10).

As an example for such a non‐epistemic consensus, Beatty (2006, 55‐64) discusses the U.S. National Academy of Sciences report entitled ‘The Biological Effects of

Atomic Radiation’ (1956). This report was a consensus statement drafted by the leading geneticists of the time, which, so Beatty argues, masked major disagreements within them about the ranges of atomic radiation that are safe to humans. Beatty identifies two reasons for masking the disagreements in this case. The first was geneticists’ worry that if they did not present a unified front, physicists would claim expertise of questions of radiation safety. The second was their shared belief that science should speak in a uniform reassuring voice to the public, because the public was not mature enough to cope with a situation in which there is no scientific consensus.

Another form of non‐epistemic consensus is a consensus around vague terms that are susceptible to multiple interpretations by different groups. For example, van der Sluijs et al. (1998) argue that there has been a stable consensus among climate scientists about the estimate of 1.5°C to 4.5°C for climate sensitivity that has survived many changes in climate theories and models. They explain this stable consensus inter alia by some ambiguity about the exact meaning of the range and the term ‘climate sensitivity’ that allows different scientists to accommodate it differently in their work, and the need to maintain a stable and credible range for question of public policy.

‐ 105 ‐

Another example is the consensus document of the United Nations International

Conference on Population and Development (ICPD) drafted in Cairo in 1994. Halfon characterizes this consensus as ‘a functional consensus’ that allows ‘disparate communities of people to act “as if” cognitive agreement were in place’ (2006, 788). He argues that this functional consensus has been achieved mainly by standardizing demographic data across many countries while maintaining interpretive flexibility of the meaning of terms such as ‘women’s empowerment’, ‘family planning’ and

‘reproductive rights’ (2006, 791‐801). If this is indeed the case, the ICPD consensus is not knowledge‐based.

To sup up the discussion so far, when a consensus exists, but veritic luck, epistemic misfortune, and non‐epistemic consensus are not present, knowledge is. For example, suppose that a group of researchers of different backgrounds who all employ different methods of inquiry reaches an agreement on a certain matter. Then we would tend to assume that their agreement is neither an accidental fluke (veritic luck) nor were they systematically or deliberately mislead in their inquiry (epistemic misfortune). In such a case, we would tend to think that they have successfully managed to reach the truth on this matter and gain knowledge. This is the basic intuition underpinning my argument.

In order to know that such a situation of knowledge obtains, we need to eliminate veritic luck, epistemic misfortune and non‐epistemic reasons as possible good explanations of the consensus. This brings me to the issue of inference to the best explanation (IBE). Under the IBE model, we legitimately infer the truth of our best explanation of a given fact. For example, if I hear squeaky sounds in my kitchen, the

‐ 106 ‐

cheese in my kitchen cupboard is chewed, and I see a mouse hole under the cupboard, I legitimately infer that there is a mouse in my kitchen, as this is the best explanation of these facts. Under the IBE model, certain explanatory merits such as simplicity, scope and elegance determine which explanations are the best. These explanatory merits are taken as indications of the likeliness of the explanation. Explanatory considerations determine plausibility, so the best explanation is the likeliest explanation (Lipton 2004,

60‐62).

What is our justification for using IBE? Why are we justified in treating explanatory merits as likeliness indicators? Here I will side with those who regard IBE as a species of inductive inference, and argue that our justification for using IBE is empirical (Ben Menahem 1990; Day & Kincaid 1994). They argue that explanatory merits are empirical generalizations, and vary from context to context. For example, if a court infers the guilt of an accused from the fact that her fingerprints were found in the crime scene, a person matching her description was seen fleeing the crime scene, etc., this inference reflects the court’s knowledge of how crimes are usually committed.

Generally speaking, we observe regularities in the world and see that some events are more frequent than others. We develop our explanatory merits based on these observations, and this is why we are allowed to make plausibility judgments based on them (Ben Menahem 1990, 322‐34). On this view, IBE does not name a formal inference, but names an abstract pattern whose force and success depend on the specific background assumptions involved, and whose specific form changes from one context to the next (Day & Kincaid 1994, 282).69

69 There are other justifications for using IBE, which I do not adopt. One view, which Ben Menahem (1990, 326‐7) attributes to Hempel and of which she is critical, justifies the use of IBE by appealing to

‐ 107 ‐

As Ben Menahem notes, in non‐philosophical contexts the disputing parties hardly ever question the connection between explanatory merits and credibility.

Rather, they differ over the question of which of the offered explanations is the best. For example, in a legal context, the accused’s defence attorney will try to raise doubts as to his client’s guilt being the best explanation of the crime, or at least to construct other plausible explanations, but he will not try to doubt IBE as such (Ben Menahem 1990,

325). Put more generally, it is not required to present an argument to justify induction in order to make a good (or a bad) inductive argument, nor is it necessary to do that to judge whether a given inductive argument is good or bad (cf. Lipton 2004, 192). I will therefore take my first claim that a consensus is knowledge based when knowledge is the best explanation of the consensus to be unproblematic and uncontroversial. What is problematic and disputable, and hence the focus of this chapter, is when knowledge is the best explanation of a consensus.

Often, we choose the best explanation by eliminating other competing possible explanations, which have less explanatory power, until we are left with one candidate

(Lipton 2004, 72‐82). When we have weeded out one explanation among competing ones, we face a dilemma: either we accept the explanation or regard the fact to be explained as some kind of a miracle (Lipton 2004, 185‐194). My own strategy in this paper is similar. I provide a sort of ‘no‐miracle’ argument for consensus. The three conditions I identify aim to rule out situations in which veritic luck and epistemic

certain formal‐structural features of good explanations. Another justification, common in the discussion of mathematical Platonism, is the so‐called indispensability argument associated with Quine and Putnam. It states, roughly, that if certain claims are indispensable for the purpose of a certain explanation, and for the purpose of this explanation we treat them as true, then for the sake of the consistency of our behaviour, we ought also to believe that they are true. For a discussion of the indefensibility argument, see Colyvan (2008).

‐ 108 ‐

misfortune are possible explanations of a consensus. When these conditions obtain, we can either attribute the consensus to be explained to shared knowledge, or regard it as a miracle of some sort.

Last, my account is explicitly fallibilistic. Because inference to the best explanation, like any other inductive inference, is a fallible inference method, my account does not guarantee that in cases in which my conditions obtain the consensus will always be knowledge‐based. Rather, it suggests that when such conditions obtain, this is very likely.

In this section I have presented a conception of knowledge as a true non‐lucky belief, which is compatible with the major theories of knowledge and captures their essence and common denominator. I have also clarified the role IBE plays in my argument. This concludes the first part of my argument according to which a consensus is knowledge based when knowledge is the best explanation of the consensus. This allows me to make the major part of my argument, which is specifying the conditions under which knowledge is the best explanation of a consensus. Before that, however, I need to clarify and defend my assumptions about idealizations and rationality, which I will do in the next section.

2. Consensus, Rationality and Idealization

Before I delve into the details of my proposed account, let me distinguish it from other discussions of consensus in philosophy. Philosophers have often suggested and discussed various idealized models of consensus among rational agents converging on the truth (Peirce 1877; Lehrer & Wagner 1981; Habermas 1984). Habermas and Peirce, at least by some interpretations, define knowledge as whatever an ideal epistemic

‐ 109 ‐

community comes eventually to agree on. While such models are interesting, it is not obvious when and how the insights they provide can be extrapolated to deal with real, as opposed to ideal, instances of consensus.

An example will illustrate this. Kelly (2005) assumes a model of equally rational agents who are all equally competent and have access to the same evidence. He argues that while they may initially have disagreements because of each individual’s limited cognitive ability, once they engage in critical deliberation, a consensus should emerge.

Kelly’s claim seems plausible, but inapplicable to typical real cases in which people differ in their competence and expertise and have only partial and non‐overlapping access to evidence. It seems that his conclusion that the existence of disagreement by itself carries no epistemic weight is, in some way, already contained in the idealized assumption that all agents are equally competent, equally rational and have equal access to the evidence. Similar considerations have led Christensen (2009, 765) to doubt the value of the literature about peer disagreement, which focuses on idealized models that are short of the difference‐making complexities of real‐life cases, to actual cases of disagreement in the real world (cf. Tucker 2003, 502‐4; Enoch 2009, 6‐7).70

In this chapter, then, I will refrain as much as possible from making idealized assumptions about agents’ rationality, competence and access to evidence. I assume that agents are mostly rational at best, biased in different ways, have different social interests, have only partial and non‐overlapping accessibility to the evidence, etc. In particular, my account pays special attention to factors such as social interests and

70 Christensen (2009, 765) raises two specific worries. The first is the difficulty with assessing credentials and recognizing reliable experts. I have discussed this worry in Chapter 1. The second worry deals with the numbers of people on different sides of the issue, and the types and degrees of causal dependence of some people’s opinions on other peopled. I discuss this worry in section 3 of this chapter.

‐ 110 ‐

social inequalities that tend to complicate things. In particular, I assume that scientists are ordinary human beings, namely mostly rational at best. My model, then, is realistic and makes as few assumptions as possible about individual agents’ rationality.

This concludes the discussion of my aims, methodology and assumptions. I will now turn to the substantive discussion of the conditions under which knowledge is the best explanation of a consensus, starting with the first condition – apparent consilience of evidence.

3. The Apparent Consilience of Evidence Condition

There is a common intuition, which states, roughly, that it is likely that a consensus is knowledge based when the parties to it are independent of one another in the ways and methods by which they arrived at their beliefs. I will explore this intuition and argue that it amounts to two separate conditions. The first, to which this section will be devoted, is apparent consilience of evidence. The second, which is social diversity, will be discussed in section 4.

3.1. First Approximation: Goldman’s Causal Independence Condition

Goldman discusses the relations between knowledge and consensus in the context of the question of whom a layperson should trust when experts disagree. He asks: ‘If a putative expert’s opinion is joined by the consensual agreement of other putative experts, how much warrant does this give a hearer for trusting the original position?’

(2001, 98). Goldman argues that at first blush, trusting the majority of experts in the community simply because it is the majority seems like a good idea, but it is actually not generally so. He gives an example of a group of people who blindly follow a guru or

‐ 111 ‐

influential opinion‐makers. In such cases, the fact that more people believe the same is not epistemologically significant.

Goldman argues that the fact that a person X is agreeing with a person Y carries weight only inasmuch as X is more likely to agree with Y that p if p is true than if p is false. Goldman argues that this is the case when X’s and Y’s routes to believing that p are at least partly causally independent. For example, when X’s and Y’s beliefs are based on independent experiments or eye witnessing, they are causally independent. When X comes to believe that p after he hears Y saying that p and critically reflects on it, X forms his belief in a partly causally independent fashion (Goldman 2001, 99‐102). For

Goldman, then, consensus is epistemically significant only inasmuch as the individuals who share it have come to the same belief in a causally independent way.71

Goldman’s view has an intuitive appeal, but there are difficulties with it. Coady argues that sometimes it makes sense to defer to the majority merely because of its size:

Most meteorologists believe that [global warming] is caused by human activity, a small minority disagree. As things stand, many people, including myself, have little but this bare fact to go on when deciding what to believe; nonetheless, I think we are justified in agreeing with the larger group of experts, just because it is larger (Coady 2006, 76).

Coady provide the following argument to justify this claim. He argues that even if Y forms his belief that p totally based on X’s view that p, this has epistemic significance if

Y has good reasons to recognize X as a particularly trustworthy within the domain of

71 Goldman’s reasoning echoes with Condorcet’s jury theorems, which state, roughly, that a sufficiently large group of individuals in which there is a sufficiently large subgroup of individuals who have a higher than chance probability to form a correct belief on a given matter will reach the reach the correct decision on that matter by the method of majority voting. The application of these mathematical theorems to concrete real‐world cases, however, is far from trivial. It is not clear in which concrete cases we would expect the conditions of statistical independence and higher than chance probabilities to obtain, or even how to judge whether they obtain in some cases. See Vermeule (2009, 28‐33) for a detailed discussion of the difficulties with applying Condercet’s theorems to concrete cases.

‐ 112 ‐

expertise to which p falls. By a layperson’s iteratively updating his degree of belief in a

Bayesian fashion based on recognizing trustworthy experts within a given domain, the fact that more experts believe that p than not, does make their view more credible

(Coady 2006, 71‐76).

There are additional reasons to think that mere numbers may make a difference.

As we have seen, it is common wisdom among philosophers that critical deliberation between individuals makes the final conclusion they reach more warranted than the conclusion at which each of them would have arrived alone. This is even if they all came to the same belief based on the same reasons, namely even if they did not form their beliefs in a causally independent way. Moreover, large groups of people who collaborate can effectively divide the cognitive labour between members of the group and reach more warranted results much faster than individuals working in isolation. Additionally, large groups do much better than small groups (Thagard 1997, 251). Although members of such groups rely on each other’s results, thus they make their respective beliefs causally interdependent, we have good reasons to trust their consensual belief, and to trust large groups more than small groups.72

Goldman’s formulation of the causal independence condition, then, is unsatisfactory. A better formulation requires a closer look at the social dynamics of consensus forming and the epistemic justificatory practices that are involved in it. In the next subsection I will present Tucker’s ‘knowledge hypothesis’ which tries to identify social conditions under which a consensus is knowledge based. I will argue that

Tucker’s theory manages to overcome the difficulties with Goldman’s formulation, but

72 A second difficulty with Goldman’s argument is that if people are causally isolated from each other, they may all reach the same conclusion for different and incoherent reasons. I discuss this problem in detail in section 5.

‐ 113 ‐

as I will argue in subsection 3.3, it does not manage to overcome the challenges that are posed by Solomon’s Social Empiricism. This will allow me to develop my alternative.

3.2. Second Approximation: Tucker’s Epistemic Theory of Consensus

Tucker (2003) identifies three conditions under which shared knowledge is the best explanation of a consensus. The first condition is that the consensus is uncoerced; the second is that it is uniquely heterogeneous and the third that it is sufficiently large. Let us look at each of them in turn.

Tucker argues that when a consensus in coerced, namely when a group of people is threatened, intimidated or bullied to hold a specific belief, their shared belief is not knowledge. An example is the scientific consensus in the Soviet Union on Lysenko’s theory of biological adaptation; scientists who did not support it were at risk of being sent to the Gulag.

Tucker talks about genuine coercion. He is not too worried about less aggressive forms of influence such as peer pressure, charisma or indoctrination. One may think that this condition is too weak and that we should be worried about such factors as well.

Tucker argues that while such forms of influence have some effect, they are insufficient for forming a consensus because with the absence of genuine coercion, there will be some dissenting people who will speak up their mind, if they believe they should

(Tucker 2003, 505). Why is that so? It seems that the second and third conditions are supposed to guarantee the existence of dissenting voices in appropriate epistemic circumstances. Let us discuss them, then.

Tucker’s second and main suggested condition is that the consensus is uniquely heterogeneous, namely no subgroup of the epistemic community shares an extraneous

‐ 114 ‐

property that may otherwise explain the agreement within it. Tucker draws an analogy between uniquely heterogeneous consensus and a controlled scientific experiment. In a controlled experiment, if members of a test group do not share any property other than the one being tested, if an effect is observed, if may be concluded that the tested property is responsible for it. Similarly, when there is a consensus in a group of people who do not all share a property such a mutual power relationship, joint interest, shared ideology or biases, and the like, the consensus may be attributed to knowledge (Tucker

2003, 506). This is why a third condition, that the group is sufficiently large, is required.

While an accidental consensus in a small uniquely heterogeneous group, so Tucker argues, is likely, it is unlikely in a large heterogeneous group.

How is Tucker’s unique heterogeneity condition an improvement of Goldman’s causal independence condition? The unique heterogeneity condition may be seen as an application of Goldman’s causal independence intuition to the group level, in a way that solves some of the difficulties with the Goldman’s individualistic formulation of it.

Rather than looking at each expert individually and see to what extent she formed her belief in a causally independent way of other experts, we look at the level of the larger epistemic community and see to what extent each subgroup formed its beliefs in a causally independent way. Each subgroup may correspond to a discipline or sub‐ discipline, to a school of though and the like, which is often united by shared metaphysical commitments and exemplars.

This formulation of the condition allows us to take into consideration the size of the group as well as the causal dependence of the parties to the consensus. When we examine subgroups rather than individuals, we may give more credence to the joint

‐ 115 ‐

view of a subgroup than we would give to the view of an individual working in isolation.

We may also give more credence to large groups than small groups. Thus, the unique heterogeneity condition manages to capture the intuition behind Goldman’s causal independence condition, while avoiding some of the difficulties with it.

Note that Tucker’s attribution of a consensus to knowledge is an IBE, more precisely a ‘no‐miracle’ argument of a sort. It would be a miracle or something like it, so

Tucker’s argument goes, if people who did not share any common property formed a consensus unless they shared knowledge. But is it indeed that unlikely for members of a sufficiently large uniquely group to form a consensus even if they do not share knowledge? As opposed to Tucker, Solomon does not think so. I the next section, I will present Solomon’s rival theory of consensus, according to which accidental aggregation of views toward a consensus that is not knowledge‐based is plausible and not unlikely. I will present Tucker’s criticism of Solomon’s theory and argue that it is unsuccessful.

Drawing on the lessons from this discussion, I will develop my own IBE account of consensus.

3.3. A Challenge for Tucker: Solomon’s Accidental Aggregation Theory of

Consensus

In Social Empiricism (2001) Solomon develops a normative account of division of cognitive labour and knowledge in science. She conceptualizes cognitive diversity in terms factors (‘decision vectors’) that influence the thinking and theory choice of individuals and communities. She identifies many types of decision vectors that affect scientists’ theory choice, such as motivational, social, cognitive, religious, and ideological decision vectors. She distinguishes between empirical and non‐empirical

‐ 116 ‐

decision vectors. Empirical decision vectors are all the factors that make scientists adopt or prefer theories with empirical success (Solomon 2001, 51‐63), where empirical success is any success of the theory that is contingent on how the world is

(Solomon 2001, 17), such as predictive success, retrodictive success, experimental success, explanatory success, and technological success (Solomon 2001, 21‐22).

Non‐empirical decision vectors are other reasons or causes for theory choice, such as preference for theories with hierarchical ideology, preference for simpler theories, or holding onto a theory because of pride (Solomon 2001, 51‐63). She argues that three conditions are necessary for an epistemically normatively appropriate dissent:

1. Theories on which there is dissent should each have associated empirical

success.

2. Empirical decision vectors should be equitably distributed, i.e. in proportion

to their empirical success. For instance, if some theory has some

technological success, a proportional number of scientists in the relevant

community should be drawn to it because of this success. Such a situation is

justified, even if these scientists are not individually rational in their

preference, for example if each of them is overly impressed by theory’s

technological success and overlooks its downsides. What matters is that at

the level of the community the theory enjoys support that is proportional to

its empirical success and due to its empirical success.

3. Non‐empirical decision vectors should be equally distributed, i.e. the same

number for each theory (Solomon 2001, 117‐120; 1994, 336‐9).

‐ 117 ‐

How does this apply to consensus? It follows from Solomon’s conditions that a consensus is justified if and only if one theory has all of the empirical success. In such a case, every distribution of non‐empirical decision vectors is OK. Such cases are typically rare. Thus, permanent dissent is typically desirable. Dissent is not a temporary glitch to be eventually overcome, and consensus is not the ultimate end of inquiry (Solomon

2001, 117‐20).

Solomon argues that the history of science shows that consensuses emerge out of an accidental aggregation of non­empirical decision vectors.73 Thus, there may be cases in which a certain theory enjoyed a consensus, for example because of the support of an influential scientist, although there were rival theories with empirical success

(Solomon 2001, 121‐35). Such consensuses are not justified. Therefore for Solomon, a consensus, even if it is uncoerced, uniquely heterogeneous and sufficiently large – as

Tucker requires – is not necessarily knowledge‐based.

Tucker rejects the accidental aggregation hypothesis as a good ceteris paribus explanation of a consensus. He interprets Solomon as arguing that consensus is likely even given conflicting interests. Or formally, if C stands for a consensus, K stands for knowledge and B1…Bn stand for biases, then ceteris paribus, Solomon argues that it is likely that

74 P(C|B1)× P(C|B2)×…× P(C|Bn) > P(C|K).

73 It is important to note that unlike Lewenstein, for example, who argues that an individual scientist forms her decision in an accidental fashion ‘based on completely contingent factors’ … ‘from salience and importance through time of day and state of hunger’ (Lewenstein 1995, 415), under Solomon’s model the accidental aggregation happens at the level of the group. It is consistent with the view, for which I have argued in the previous chapter, that individual scientists form their judgment in a consistent way. 74 When we examine the question of whether cause X is the best explanation of effect E, we are interested in the question of how likely E is given X. Since we are discussing here the question of

‐ 118 ‐

In other words, Tucker takes Solomon to argue that ceteris paribus, if we divide a group of people into n subgroups, where each subgroup shares a common bias,75 the probability that a subgroup that shares a bias B1 and a subgroup that shares a bias B2, and so on, would all reach a consensus is greater than the probability that the group would reach a consensus if the consensus view amounted to knowledge.

Against Solomon, Tucker argues that for Solomon’s claim to be true, we need to assume that:

(a) the likelihood for the consensus given each bias is initially very high;

(b) the probability of the consensus given shared knowledge is very low.

Tucker takes these two assumptions to be implausible, thus he rejects Solomon’s accidental aggregation hypothesis (Tucker 2003, 503).

There are several difficulties, however, with Tucker’s objection. As for (a),

Tucker is wrong to assume that the likelihood of consensus given each bias is very low.

It seems plausible that sometimes biases will all pull at the same direction. Consider the following example. In 2005, leaders of all the major religions in Israel – Orthodox

Judaism, Shia Islam, Roman Catholic, Greek Orthodox and Armenian Christianity – issued a joint consensual statement against having a gay pride parade in the city of

Jerusalem, which they consider holy (Goodstein & Myre 2005).76 What is unique about

when knowledge is the best explanation of a consensus, the relevant probability to consider is P(C|K) rather than P(K|C). 75 Note that this way of formulating things already seems unrealistic and might not accurately reflect Solomon’s claim, since so‐called biases are usually not mutually exclusive. For example, in a community of physics, men tend to be white. For the purpose of my argument, I am not going to delve into this point. 76 Israel’s Jewish population consists of three major groups: Orthodox religious Jews, tradition keepers, who selectively keep some of religious commandments and to a lesser extent than religious Jews, and secular Jews. The Jewish Orthodox religious establishment is authorized by law to be the sole provider state‐religious services such as marriage and divorce, Kosher certification and religious burial to all the Jewish population. Progressive branches of Judaism, which are common in North

‐ 119 ‐

this particular agreement is that despite the common ancient history of these religions, these religious groups hardly ever agree on anything. They disagree on fundamental religious truths, such as the number of true Gods and true prophets. They disagree about the correct interpretation of the Scriptures, and even about which books belong to the Scriptures. They disagree about questions of control of land, territory and religious sites in Jerusalem. They disagree about most issues in regional and local politics. But they happen to agree that there should not be a gay pride parade in

Jerusalem. When we look at the vast issues on which these groups usually disagree, it becomes clear that the joint statement represents an ad­hoc coalition, and nothing more. On its own, the fact that representatives of the major religions of Jerusalem happen to agree on something does not carry any epistemic weight over and above the reasons they use to justify their claim.

Are such ad hoc coalitions so unlikely? The answer is negative. When people and groups have a variety of interests and carry views on a variety of issues, they are most likely bound to agree once in while on some things. We may sometimes be surprised that an ad­hoc agreement was formed on a particular matter, but we should generally expect such agreements to be formed on occasion. Different groups of people often have mutual interests and beliefs which will cause them to reach a consensus on a particular matter, in spite of disagreeing on many other things. If the social dynamics of science are similar to those of the rest of society, we should expect such accidental ad hoc consensuses to occur in the sciences on occasion as well.

America and would probably have not taken part in such a statement, have very little presence in Israel and no official state recognition.

‐ 120 ‐

As for (b), the prior probability of the consensus given knowledge is exactly what is in dispute. Even if people share knowledge, they might not reach consensus for a variety of reasons such as mutual misunderstanding, or difficulties to communicate or transcend personal rivalries and bickering (Thagard 2000, 236‐7). The prior probability of the consensus given knowledge cannot be assumed to be high without begging the question.

In response to these two claims, Tucker may still argue that in the inequality statement above, when n is large enough, the left side (‘biases’) still goes to zero. In other words, so this argument goes, even if the probability of consensus given each bias is initially high and the prior probability of consensus given knowledge is low, when there are enough biases, a convergence of all of them toward a consensus is still very unlikely.

This claim, however, is problematic because the expression on the left hand side of the inequality statement is misleading. It describes the probability of complete agreement with no dissent whatsoever, which is very low anyway. For example, suppose we have a group of 100 people, and each person is .95 likely to believe that p. The probability that all 100 people will agree that p is close to zero, but the probability that about 95 of them will believe that p is close to one. In real‐life situations, we are not interested only in cases of complete agreement, but also in cases of a very wide agreement with only a small dissent, which are not captured by the left‐hand side of the inequality statement.

In fact, Solomon’s historical counterexamples – all from the twentieth century – of accidentally formed and epistemically unjustified consensuses are of wide agreement

‐ 121 ‐

with only minor dissent, rather than complete agreement. They include the consensus on the Central Dogma in Biology, the Male Variability Hypothesis, the Extracranial‐

Intercranial Bypass surgical practice, the Ovulation theory of Menstruation, and the excess acidity causal theory of peptic ulcers (Solomon 2001, Ch. 7).

Tucker argues that these examples do not refute his claim because they are not examples of a genuine consensus. If there are two sides, so Tucker argues, and there is a dissenting view, the agreement does not amount to a true unity of opinions. But in real life cases, to which Tucker’s theory purports to apply, a complete uniform agreement is hardly ever present, and we should not require it in order to be able to say that a consensus on some issue exists. Pace Tucker, the mere existence of a dissenting view is not enough for discounting certain wide agreements as genuine consensuses. For example, it is widely believed today that AIDS is caused by the HIV virus. There are, however, few dissenting scientists who disagree, and their belief is dismissed as groundless by the vast majority of the scientific community.77 The belief that HIV does in fact cause AIDS constitutes a basis for research, treatment and prevention practices.

It is taught and presented as a fact in university curricula and to the public. The dissenting beliefs are suppressed and marginalized by the mainstream scientific community, and their proponents face extreme difficulties challenging the majority belief and publishing their views. In light of that, though a pocket of dissent exists, it would still be fair to say that there is a consensus that HIV causes AIDS. Tucker’s mathematical formula that purports to describe the existence of a consensus in a community is therefore inadequate.

77 The chief advocate of this dissenting view is biologist Peter Duesberg from the University of California, Berkeley, and he argues for it in his (1996). See Epstein (1996, Ch. 3 & 4) for the history of the AIDS controversy.

‐ 122 ‐

3.4. The Apparent Consilience of Evidence Condition

So far I have argued that although Tucker’s account is an improvement to Goldman’s causal independence condition, it does not successfully manage to refute Solomon’s accidental aggregation hypothesis. Does this mean that an IBE ‘no‐miracles’ account of knowledge‐base consensus is hopeless? Not necessarily. It is still intuitively plausible that all things being equal, when many people who do not share much in common come to an agreement, shared knowledge is the best explanation of their consensus. What this shows is that Tucker’s criteria fail to adequately capture this ‘no‐miracles’ intuition. Let us, then, examine what underpins this intuition.

A closer look at Tucker’s and Goldman’s theories reveals that they are not interested in social unique heterogeneity and causal independence as such. Note that if under Goldman’s model, different individuals happened to form their belief in exactly the same way despite their causal independence, then their combined consensual view would carry no more epistemic weight than the view of each individual alone. Goldman and Tucker take such a scenario to be very unlikely. Their respective conditions of causal independence and unique social heterogeneity are proxy conditions. Ultimately, what matters for them in their respective accounts is that different agents form their beliefs in different ways, and it is assumed that causally independent people or subgroups of people who do not have an extraneous property in common will tend to form their beliefs in significantly different ways.

Underlying their accounts is the thought that the more varied the ways people form the same belief, the more support it has. If many people come to believe that p in

‐ 123 ‐

significantly different ways, this is probably because that p is true or approximately true. Knowledge would be the best explanation of why they all have the same belief.

This is, in fact, an application of the general notion of robustness to the social context. Robustness is the idea that ‘hypotheses are better supported with plenty of evidence generated by multiple techniques that rely on different background assumptions’ (Stegenga, 2009, 650).78 Producing evidence using multiple techniques and under different background assumptions aims at eliminating influences that are accidental to the particular way a hypothesis is tested. For example, if the same pattern is observed when the same sample is placed in different types of microscopes that operate on different principles, it is likely that the observed pattern is accurate, rather than by‐product of the particular way a certain microscope operates.

In the social context, the robustness principle would be that when a consensus is built on an array of evidence that is drawn from a variety of techniques and methods, it is less likely to be an accidental by‐product of one technique — and all the more likely to be knowledge‐based. I suggest that in the social context, when we talk about convergence multiple techniques, we mean apparent consilience79 of different types of evidence. For example, if both animal studies and human studies support the same conclusion, we would tend to think it is more likely to be true.

78 Stegenga (2009) argues that an intuitive notion of robustness that does not specify what robustness would mean in concrete cases in which there are many lines of evidence is of little value and undercuts the integrity of the very notion itself. Stegenga’s point is well taken, and my account may be seen as specifying the conditions for a robust consensus, bearing in mind, though that my account specifies only sufficient conditions for a consensus being knowledge‐based. 79 I borrow the term ‘consilience’ from William Whewell, who coined it. Whewell talks about the principle of ‘consilience of inductions’, according to which hypotheses are more supported when they independently stem from different inductive inferences (1858, 87‐90).

‐ 124 ‐

Apparent consilience of evidence is what Tucker’s unique heterogeneity condition ultimately strives for. How is that so? Typically, different groups of people – different disciplines and sub‐disciplines – use different methods and different types of evidence. For example, in the Bendectin case study I will discuss in the next chapter, evidence was used from animal studies and from human epidemiological studies. Both of them constitute different types of evidence that correspond to localized groups – epidemiologists and toxicologists – that have their own journals, societies, etc.

Not only do disciplinary boundaries make researchers use different types of evidence, but so do geographic, national and other social barriers. For example, historians of science talk about ‘national styles’ in doing science. A famous example is the divide in the nineteenth and early twentieth centuries between French and English physicists and chemists, where French scientists strongly favoured abstract mathematical reasoning, and English scientists strongly favoured reasoning based on concrete mechanical models and visual diagrams (Nye 1993). Because in reality, different social groups tend to use different types of evidence, Tucker’s unique heterogeneity condition is a good proxy for apparent consilience of evidence.

While unique heterogeneity is a good proxy for apparent consilience of evidence , it is not good enough. It is not always the case that different social groups and even different disciplines use different methods and different evidence of different types. It follows from the argument in the last subsection that we cannot assume it will be so unlikely to find a uniquely heterogeneous group whose members happen to use the same methods and rely on the same evidence. We should therefore explicitly specify

‐ 125 ‐

that we are interested in apparent consilience of evidence, rather than unique heterogeneity, as a social epistemic indicator for a consensus being knowledge‐based.

At this point one may wonder why I have termed the condition apparent consilience of evidence, rather than simply consilience of evidence. As I will argue in

Chapter 5, which concerns the role of social values in evidential reasoning, very often determining whether a given body of evidence supports a theory and to what degree is not a straightforward matter. The judgement that individuals and communities make about evidential support is influenced by social values, and may change over time. In particular, as I will argue, social values affect the process of combining different lines of evidence, and the judgment people make about whether lines of evidence converge.

The condition of apparent consilience of evidence is weaker than actual consilience of evidence. It only requires that all existing evidence seem to support the consensual view. The apparent consilience of evidence condition will become clearer in the next chapter, where I review the Bendectin case‐study and argue that the consensus in that case fails to meet it. In a nutshell, in the Bendectin case‐study, the condition is not met, because the scientific community has not pursued research in animal studies, in spite of the fact that some evidence that was produced from animal studies seemed not to lend support to consensual view that has emerged.

If my model required actual consilience of evidence rather than just apparent consilience of evidence it might run the risk of becoming trivial. Under leading theories of epistemic justification, such as evidentialism and coherentism, when all evidence supports a belief, it is epistemically justified. Thus, if my account required actual consilience of evidence, it might amount to the triviality that a consensus is likely to be

‐ 126 ‐

knowledge‐based when the consensual belief is epistemically justified. There is not much point in arguing that. By contrast, the suggested condition only requires that it does not seem at a certain point in time that there is any known evidence that is being ignored, overlooked or otherwise suppressed by the members of the consensus group.

4. The Social Diversity Condition

In the last section I have argued that the apparent consilience of evidence condition captures the robustness intuition applied to the question of when a consensus is knowledge based better than Goldman’s causal independence and Tucker’s unique heterogeneity conditions. As you recall, robustness is the notion that hypotheses are better supported with plenty of evidence generated by multiple techniques that rely on different background assumptions. The apparent consilience of evidence condition obviously addresses the ‘multiple techniques’ part of the robustness definition, but what about the ‘different background conditions’ part of it? Does the consilience condition, on its own, adequately capture it as well? In this section I argue that the answer is negative, and to fully capture the robustness intuition, social diversity needs to be added as a separate condition.

In the social context, the multiple background condition is often equated with social diversity, as members of different groups tend to have different background assumptions. With respect to consensus, Longino writes that a ‘diversity of perspectives is necessary for vigorous and epistemically effective critical discourse […] When consensus exists, it must be the result […] of critical dialogue in which all relevant perspectives are represented’ (Longino 2002, 131).80

80 While I agree with Longino on this point, as I will show in section 6, my own account differs from

‐ 127 ‐

Diversity has many epistemic benefits. Diversity is more likely to generate new research questions, identify limitations with existing models and propose new models, propose a fuller range of alternative hypotheses and interpretations of data, access more accurate and complete data from human subjects, open up new lines of evidence, reveal ‘loaded’ language in descriptions of phenomena, and more adequately identify and weigh potential risks (Intemann 2009).

Mill argues that a person’s beliefs are a product of ‘his party, his sect, his church, his class of society […] his own country or his own age’ (Mill 1859/1993, 86). Feminist epistemologists argue that because certain perspectives are often inseparable from certain social identities, even in open and critical settings, there is a limit to people’s ability to transcend their original background and free themselves of their biases and prejudice.

Thus, we need actual social diversity, not just the negative freedom to raise criticism

(Fricker 2007; Okruhlik 1998). In the same spirit, Longino argues that the absence of women and ethnic minorities from a scientific consensus, even if it not intentional, constitutes a serious cognitive flaw, ‘as their absence reduces the critical resources of the community’ (Longino 2002, 132). If we are to take seriously the fact that actual people are far from perfect rational agents, and that it is very difficult for them to get rid of all the social influences on their beliefs that stem from their specific social background, then de facto diversity is necessary condition for a consensus to be knowledge‐based. Diversity has an indispensable epistemic role.

One might object and argue that the condition of social diversity is too strong.

Women, for example, have been historically underrepresented and excluded from the process of generating much of our current scientific knowledge. This shows, so this hers in important ways.

‐ 128 ‐

objection goes, that while diversity may be epistemically beneficial, it not necessary.

Diversity, however, comes in degrees. It is not ‘all or nothing’. The exclusion of women from certain fields does not mean lack of diversity, but less diversity.

Nevertheless, lack of sufficient diversity is indeed an epistemic problem, and theories that have been produced without sufficient diversity may indeed deserve less warrant than they currently enjoy. The prevalence of sexist and other culturally problematic background assumptions in science has been extensively studied by historians and philosophers of science. I will mention a few prominent examples. Martin

(1991) argues that because researchers have been blinded by a gendered stereotypical

Sleeping Beauty/Prince Charming model of the egg and the sperm, they have overlooked the active part that the egg plays in the process of fertilization. In cultural anthropology, Longino and Doell (1983) argue that social stereotypes about man as inventive and resourceful and woman as passive and submissive contributed to the development of ‘man the hunter’ theory of human cultural development, where alterative theories that fit the same empirical evidence and attribute significant positive contribution to women were not seriously considered. Gould (1996, 240) shows how in the 1920s, intelligence testing conducted by the US government on newly arrived immigrants to the US presupposed knowledge that was unquestionably specific to

American culture. Keller (1983) argues that biologist Barbara McClintock’s unique methods and theories in heredity, which were initially unrecognized but ultimately won her the Nobel Prize, stem partly from her being a woman with a different background from other researchers in her field. As Okruhlik (1998) argues, the important thing about such examples is not so much which theories ultimately proved to be true, but

‐ 129 ‐

that the flaws in established orthodox theories and possible alternatives to them have not been seriously considered due to lack of sufficient social diversity in the scientific community.

One may argue that such examples only show that diversity is needed in the special sciences, namely the biological and social sciences, but not in the hard sciences, such as physics and mathematics. This objection, however, is problematic. First, though the examples above are from the biological and social sciences, they do not all directly address questions of gender and society. McClintock, for example, studied heredity in maize rather than humans. Second, the influence of social stereotypes is not restricted to the special sciences. For example, Wagner (2009) argues that the use of gendered language in the formulation and proof of the mathematical theorem known as the

‘stable marriage theorem’ has blinded mathematicians from some of the potential mathematical implications of the theorem. While I do not want to overstate the case and claim that mathematics is male‐biased, it seems that the burden of proof lies with those who want to argue that diversity only matters in special sciences. After all, in the above mentioned examples, the theories in question were thought to be objective and value‐ free until the cultural perceptions embedded in them were pointed out. Today, women are still underrepresented in the hard sciences, thus the potential epistemic benefits of more diversity in these sciences is yet to be discovered.

One may argue that while social diversity is important, the condition of social diversity is reducible to the first condition of apparent consilience of evidence.

Ultimately we are interested in realizing the robustness condition for consensus.

Therefore, while diversity may be instrumental in bringing about a variety of

‐ 130 ‐

background assumptions required for robustness, it is not necessary. It is sufficient, so this objection goes, that the different converging lines of evidence that the apparent consilience of evidence condition requires seem to be also based on sufficiently varied background assumptions, where social diversity is just one way of ensuring that.

Against this objection I argue that social diversity is required as a stand alone condition that is irreducible to apparent consilience of evidence. This is because the production and assessment of evidence is influenced by social factors. Therefore, the appearance of consilience of evidence may itself be attributed to some underlying non‐ epistemic social reality, rather than to the existence of knowledge in the community that shares the consensus.

I will give a detailed account of the relations between social values and evidence in chapter 5, but for now, consider the following thought experiment. Suppose you find out that a scientific consensus exists, according to which passive smoking does not raise the chances of lung cancer. Suppose that this consensus exhibits apparent consilience of different lines of evidence. Studies of different types support this conclusion: human epidemiological studies show no significant statistical correlation between passive smoking and lung cancer, structural‐analysis studies suggest that cigarette smoke undergoes some chemical reaction in the open air that reduces its carcinogenic effects, and so on. Suppose further that the different studies do not seem to be based on some common problematic background assumptions. This consensus, so the objection goes, is likely to be knowledge based.

Suppose that you later find out that all parties to the consensus have some financial ties with the tobacco industry, and that all the studies were supported or

‐ 131 ‐

partly supported directly or indirectly by tobacco companies. Is knowledge still the best explanation of the consensus? Not any more. Regardless of what you thought of the consensus and the evidence before, they now become suspect. A better explanation of the consensus may be that the tobacco industry is responsible for bringing it about. For example, it may have given some leading researchers sufficient incentives to produce evidence that supports its interests and make the evidence look convincing enough for other members of the scientific community to form a consensus.

The upshot of this thought experiment is that it is not enough for a consensus about the harmlessness of passive smoking to exhibit apparent consilience of evidence.

Rather, it must be also socially diverse, namely shared by researchers from both the private and public sectors, with different financial ties, smokers and non‐smokers, and so on. Hence, social diversity is irreducible to apparent consilience of evidence.

In this section, I have argued that social diversity is an independent condition that is required to ensure that a consensus be based on sufficiently varied background assumptions to be likely to be knowledge based. There is, however, a potential problem with a consensus that is based on varied background assumptions, which is that they may be incompatible with one another. In such a case, the consensus may be based on inconsistent reasoning, and hence may not be knowledge‐based. I address this worry in the next section, where I discuss the meta‐agreement condition.

5. The Meta­Agreement Condition

Are social diversity and apparent consilience evidence sufficient indicators for knowledge being the best explanation of a particular consensus? A worry that arises is that if people who are causally isolated or have different background assumptions form

‐ 132 ‐

a consensus, it may not be knowledge‐based, as it may be based on incoherent reasoning.

In order to overcome this worry, we need to formulate a meta‐agreement condition. Meta‐agreement means that people agree regarding what it is they agree upon and share the same fundamental background assumptions. It ensures that the agreement will be genuine and not superficial. In this section, I argue that a joint commitment to using the same evidential standards and a joint acceptance of the same formalism and ontological schemes, even if members to the consensus differ in their conceptual interpretation of these schemes, is sufficient to meet the meta‐agreement condition for knowledge‐based consensus.

Drawing on Kuhn, we can identify three types of possible objects for a consensual meta‐agreement. The first is shared formalism (or in Kuhn’s terms, ‘symbolic generalizations’), for example, f=ma in Newtonian physics and 2H2O → 2H2 + O2 in modern analytic chemistry (Kuhn 1970, 182‐3). The second type of objects is ontological schemes (‘metaphysical models’) which are descriptions or models of the building blocks and furniture of the world, such as that matter is composed of particles.

As Kuhn notes – a point which will be of significance later on – members of a group may vary in their strength of their commitment to such models along a spectrum from mere heuristics to ontological models of the world (1970, 184). The third is evidential standards (‘exemplars’)81 which are model solutions that show how to apply the formalism to solve specific problems (Kuhn 1970, 187‐91).

81 See Kusch (2002) at 152‐7 for an account of evidential standards as shared exemplars.

‐ 133 ‐

The objects of meta‐agreement are not necessarily linked to a specific Kuhnian disciplinary matrix. They may also be part of a certain method of reasoning. As Hacking argues, statistical reasoning comes equipped with an ontological scheme by which it describes the world as composed of populations, which have properties such as distribution and standard deviation. The procedures for calculating an average or a standard deviation are defined independently of what the population represents in the world – be it people or particles (Hacking 2001, 184). Similarly, the method of scientific reasoning with an analogical model comes equipped with an ontological scheme that describes the unobservable microscopic world as analogical to macroscopic mechanical models. Here as well, scientists may vary in their commitment to the reality of the objects such ontological schemes describe.

Let us now turn to examine the worry that arises from a causally independent consensus and see whether meta‐agreement on formalism, ontological schemes and evidential standards overcomes it. This worry is raised by Fuller in his discussion of the social epistemology of consensus. Fuller draws a distinction between two types of consensus, which lie on two opposite extremes of a spectrum: essential consensus and accidental consensus. In an essential consensus, a group forms a collective decision for the same thing using shared standards of evidence and sense of relevance. In an accidental consensus, each individual forms the same belief on her own and for her own reasons. (Fuller 2002, 208‐9). Fuller writes:

‐ 134 ‐

The paradigm of an accidental consensus is the kind of agreement that pollsters find in the course of surveying public opinion. Although the pollster presumes that everyone surveyed understands the question in the same way, he usually does not check. And indeed, studies show that by paraphrasing a question one way instead of another, the extent of consensus may be manipulated. This suggests that the degree of agreements on what the question means is never particularly deep (2002, 210).

Indeed, without further checks, a random group of people surveyed cannot be said to have knowledge. As you recall from section 1, knowledge must exclude veritic luck. Accidental consensus fails to do that. As Fuller notes, it is contingent and easily manipulatable. Even if it is right, it may have easily been wrong. Thus, it cannot be knowledge‐based.

Fuller regards a consensual agreement by scientists that is restricted to shared formalism and evidential standards as an accidental consensus rather than an essential consensus. Fuller claims that ‘to establish that the scientists have agreed upon a certain mathematical formalism […] is hardly enough to show that they have decided to pursue something in common’ (2002, 219). He worries that such a consensus may mask deep conceptual differences, due to incommensurable assumptions that may surface later

(Fuller 2007, 10). For a consensus to count as an essential consensus, Fuller requires that the parties to the consensus also have uniformity of interpretation and conceptual understanding of the theory or views in question (Fuller 2002, 219).

To support this claim, Fuller provides the following argument, which I will call the argument from logical conjunction: In a case of a consensus that is restricted to shared formalism, if we look at each person in isolation and construct a logical conjunction of the person’s reasons to support the consensual view, we will most likely get a logically consistent set of beliefs. Hence, we may regard each person’s belief as

‐ 135 ‐

rational. However, because the members to consensus differ in their conceptual interpretations, if we construct a logical conjunction of all the members’ reasons to support the consensual view, we will most likely end up with an incoherent set. Hence, so this argument goes, the consensus is irrational and cannot be knowledge‐based

(Fuller 2002, 220‐1; cf. Pettit 2006).

While I agree with Fuller that only an essential consensus can be knowledge based, I disagree that a uniformity of interpretation and conceptual understanding is required for an essential consensus. I argue that an agreement on shared formalism, ontological schemes and evidential standards is sufficient to satisfy the meta‐agreement condition.

A common distinction that is drawn between belief and acceptance (van

Fraassen 1980, 69; Stalnaker 1984, 79; Cohen 1992) may help us understand the problem with Fuller’s argument from logical conjunction. For the present discussion, there are several relevant differences between the attitude of belief and the attitude of acceptance. To accept a claim is to take it for granted in one’s reasoning, and it is possible to accept a claim without believing it. Acceptance often results from a consideration of one's goals, which may be both epistemic and non‐epistemic goals, while beliefs are not deliberately acquired in order to advance goals. Acceptance is voluntary, whereas belief is involuntary. Belief results in a feeling, in particular, a feeling that something is true; acceptance involves no such feeling (Wray 2007, 340;

2001, 325).

Are views adopted by a group, such as a consensus, a species of belief or acceptance? As Wray argues, they are a species of acceptance. Unlike proper beliefs,

‐ 136 ‐

views are adopted by a plural subject as a means of realizing the group’s goals. Further, the sorts of considerations required to bring about a change in a group’s view are different from the sorts of considerations required to change a person’s beliefs. The considerations required to change a person’s belief are for the most part epistemic – they relate to things such as its truth and falsity and the evidence that supports it. By contrast, the considerations required to change a group’s view are for the most part pragmatic, and relate to the goals for which the group adopted the view. Members of a group may be persuaded that holding a different view may better promote their shared goals, or they may choose to change their goals and change the view they adopt accordingly (Wray 2001, 326‐7).

Wray’s argument is consistent with my own analysis of consensus. In section 1, I distinguished between an epistemic and non‐epistemic consensus. People form a non‐epistemic consensus in order to promote non‐epistemic goals such as fighting a common foe. A non‐epistemic consensus is clearly a species of acceptance and not of belief, hence, so is an epistemic consensus. This is because the relevant distinction between an epistemic and a non‐epistemic consensus is not that of between belief and acceptance. Rather, the relevant distinction between them concerns their status as justified and true. Put differently, when we want to know that a consensus, e.g. the scientific consensus that human‐caused global warning is occurring, is knowledge based, what we are really asking is whether it is likely to be justified and true.

There are also particular reasons to characterize a scientific consensus as a species of acceptance rather than belief. First, a scientific theory sometimes faces difficulties in explaining anomalies, but with the absence of a better theory, scientists

‐ 137 ‐

stick to it and may be justified in doing so, though they may not take it to be true.

Second, because theory is underdetermined by data, scientists sometimes choose a theory based on pragmatic considerations such as which theory is easier to work with, where these considerations are detached from the question of the truth of the theory. In such cases, acceptance, rather than belief, correctly characterizes the scientists’ attitude toward their theories, and seems also to be the justified attitude toward the theories

(Cohen 1992, 90‐2).

Last, historically, shared belief, in the sense of having the same the same beliefs about the structure of the world, is not required for pursuing science. Only reasoning about it in a certain way is required. For example, some scientists in the 18th century working within the caloric paradigm, which conceptualized heat as fluid called ‘caloric’ that was supposedly composed of unobservable particles, were not committed to the reality of caloric and adopted an instrumentalist stance toward it. Nevertheless, they all used the notion of caloric in their reasoning (Psillos 1994, 166‐8; Fox 1971, 24).

If consensus is a species of acceptance rather than belief, then Fuller’s argument from logical conjunction is misplaced. It states that the set of all the group members’ reasons for believing the consensual view is incoherent. It does not follow from that that the conjunction of their reasons for their acceptance of the consensual view is incoherent. Reasons for acceptance are pragmatic and depend on the collective goals of the group and the individual gaols of its members. They only need to be logically coherent to the extent people share the same goals. For example, in a group of scientists, it may be the case that different persons have different views about the goals of science, but as a group they agree on a modest epistemic aim such as empirical adequacy as a

‐ 138 ‐

common denominator.82 Their collective set of reasons to accept a consensus view need only be logically coherent with respect to that goal. Pace Fuller, true uniformity of conceptual beliefs is not required.

It is worth emphasizing even though an epistemic goal such as empirical adequacy may be modest, it is not trivial. It is often hard to realize it, and achieving justified agreement on an empirically adequate theory is often a non‐trivial cognitive achievement. It seems that Fuller downplays the significance of such an achievement, which, historically, has been often hard to reach. Moreover, conceptual beliefs go beyond the shared epistemic standards and ontological schemes. They are about the nature of the shared standards and schemes, thus the standards themselves cannot decide between them. Conceptual beliefs seem to belong to the private realm of the individual agents, rather than the collective public realm of the group.

Requiring meta‐agreement in the sense of joint acceptance of the fundamental evidential standards, ontological schemes, and shared formalism strikes the right balance between preventing the consensus from being accidental, in a way that disqualifies it from being knowledge‐based, and allowing the parties to the consensus to maintain a diversity of perspectives, views and interpretations, as the robustness conditions require (cf. Baigrie & Hattiangadi 1992). As opposed to Fuller’s view, then, an essential consensus can exist in spite of latent conceptual disagreements between members of the group, and it may therefore be knowledge‐based.

82 It is important to stress, though, that under my account, if two communities, such as the proponents of the caloric theory of heat the dynamic theory of heat, have incompatible ontological schemes and are not united by shared formalism, they cannot be said to agree, even if their theories are empirically equivalent.

‐ 139 ‐

While I argue that agreement on formalism and the like satisfies the meta‐ agreement condition for knowledge‐based consensus, there is an obvious caveat. When a controversy is explicitly and directly about the metaphysical interpretation of a shared formalism, such as in the case of the controversy over the interpretation of quantum mechanics (Solomon 2001, 127‐9), we should qualify the scope of the consensus and limit it only or the empirical content of the theory as expressed by the shared formalism. Similarly, when a difference in conceptual understanding makes some members of an epistemic community reject a certain evidential standards in spite of their commitment to a shared formalism, such as in the case of mathematical intuitionists’ rejection of non‐constructive proofs, we cannot attribute a knowledge‐ based consensus to the group as a whole.

This section concludes my discussion of the three conditions that jointly constitute a sufficient condition for knowledge being the best explanation of a consensus and consequently the consensus being knowledge based. In the next section I will contrast my theory with Longino’s Critical Contextual Empiricism, which attempts to answer the question of when a consensus is knowledge based by identifying certain norms to which a community must adhere, and argue for the merits of my account. .

6. A Comparison with Critical Contextual Empiricism

A rival to my account is Longino’s Critical Contextual Empiricism (CCE). CCE is a social epistemology that takes seriously the insights of the field of STS, while not dispensing with the normativity of traditional epistemology. CCE defines knowledge as a type of consensus. Under CCE, members of an epistemic community must agree on shared standards of justification (‘epistemic acceptability’) and shared criteria that determine

‐ 140 ‐

when they collectively deem a representation of a target system as sufficiently adequate

(‘conformation’). This agreement must be reached through process of critical deliberation and scrutiny that satisfies four conditions:

1. there are public venues of criticism such as professional journals and

conferences;

2. there is uptake of criticism – members of the community appropriately respond

to the criticism and revise their beliefs accordingly;

3. there are publicly recognized standards of evaluation of theories;

4. there is tempered equality of intellectual authority – intellectual capacity and

relevant expertise are the only criteria by which people are given the right to

participate in the collective discussion, and all those with intellectual capacity

and relevant expertise can in fact realize their right to participate, regardless of

gender, race, etc. (Longino 2002, 128‐140).

CCE, then, defines knowledge as a consensus that satisfies certain norms of critical discussion. I do not deny the epistemic importance of critical discussion and the norms that Longino specifies. Moreover, my account shares important features with

CCE, such as an emphasis on the epistemic importance of social diversity. But my account also differs from CCE in important ways. In this section I will argue that the four conditions that CCE specifies are neither necessary nor sufficient for a consensus being knowledge‐based. I will use CCE to demonstrate the difficulties with any account that attempts to flesh out the conditions under which a consensus is knowledge based in terms of collective norms to which members of the consensus group must adhere, and I will argue that my account successfully overcomes these difficulties.

‐ 141 ‐

As Longino admits, the four conditions she specifies are to some extent a sketch, and the fine details of the standards of critical scrutiny required for knowledge still need to be fully worked out (2002, 133‐4). Not fully worked‐out, though, they are in danger of being too permissive or too restrictive. Interpreted too leniently, a community of like‐minded people that adopts standards of critical discussion, such as a group of creationists with their own peer‐reviewed journals, may be said to satisfy to

Longino’s norms, although they may not share knowledge. Interpreted too strictly, a scientific community whose members refuse to take seriously criticism they regard as too far fetched to begin with, such as contemporary evolutionary biologists who do not engage with creationists, may be said to fail to meet Longino’s conditions and not to share knowledge (Goldman 2002b).83

Even if CCE’s exact standards of critical scrutiny are exactly formulated such that they are neither too permissive nor too restrictive, they are still neither necessary nor sufficient for knowledge. With respect to being necessary, much of our current scientific knowledge has simply not been generated by critical scrutiny of this sort. Moreover, the standards of critical scrutiny Longino requires seem too high for ordinary humans being to meet. To what extent can scientists realistically be expected to engage in an impartial and equal critical discussion while transcending all their biases and prejudice?

One may object by saying CCE does not require scientists to eliminate all social influences, but only ones that are obstructive the transformative criticism of consensus to knowledge (Wray 1999, S547). Nevertheless, this requirement still seems too unrealistic. Recall that one of Logino’s main motivations to develop CCE was to develop

83 For Longino’s response to this line of criticism see her 2002 at 158‐159.

‐ 142 ‐

an STS‐informed normative social epistemology. If there is anything the vast empirical work in STS teaches us is how difficult it is for people to be rational and transcend their biases.

As for being sufficient, there may be instances of consensus that meet CCE standards of critical scrutiny, but are not knowledge‐based. For example, Solomon and

Richardson (2005, 214) argue, contra Longino (2002, 161‐2), that witchcraft detection practices in pre‐ and early‐modern Europe satisfied CCE standards. Even if their particular historical account is wrong (Gibeault 2008), as we have seen in section 3.3 , the existence of dissent is a contingent matter. It is not that unlikely that a diverse group just happens to have a particular distribution biases that would prevent its members from raising relevant criticism. Openness to criticism does not guarantee the existence of actual relevant criticism, and in certain social circumstances, epistemically undesirable beliefs can survive for a long time although the community is in principle open to criticism. As Solomon and Richardson (2005) argue, this general lesson to draw from this argument is that conditions for knowledge cannot be formulated solely in terms of the procedures that a community should follow. They must also say something about the conditions the end product the community reach must meet. My account specifies such conditions.

There is yet another problem with CCE, which highlights a major difference between CCE and my account. While CCE regards consensus as the final aim of a process of critical deliberation by building consensus into the definition of knowledge, my account will actually tend to be suspicious of a process which explicitly or implicitly aims at consensus. This is because such a consensus is ceteris paribus less likely to be

‐ 143 ‐

knowledge‐based than a spontaneously occurring consensus (cf. Solomon 2007b). Recall that according to my account, a consensus is knowledge based when knowledge is the best explanation of it. Whether knowledge is the best explanation of a particular consensus depends on the availability of other explanations for that consensus. If a certain community was aiming at consensus, a good explanation and possibly the best explanation of why it has reached consensus is that it was aiming at consensus all along.

If this is so, it is less likely to be knowledge‐based. Since CCE requires that a community set consensus as its goal, when a community that follows CCE recommendation reaches a consensus, its best explanation may be that it was aiming for it to begin with, rather than knowledge.

Last, a serious difficulty for CCE, with which, by contrast, my account can successfully deal, is the problem of minority stubbornness and manufactured uncertainty. What happens when some minority members within the community, sometimes backed by powerful and wealthy bodies such as tobacco or pharmaceutical companies, insist that some claims have not yet satisfied the community‐defined epistemic standards and that more and more critical scrutiny is still required? When, if ever, is a community allowed to move on, and according to which standards? It seems that CCE cannot deal with cases of manufactured uncertainly, in which certain people with a vested interest against the acceptance of a certain theory cynically and deliberately keep on insisting on and more critical scrutiny, no matter what the evidence is. By contrast, my approach allows us to dismiss such bodies, as the best explanation for their ongoing insistence is their vested interest in a certain outcome, and evaluate the remaining consensus by my suggested criteria.

‐ 144 ‐

In this section I have contrasted my account of knowledge‐based consensus with

Longino’s CCE. I have argued that the IBE approach manages to successfully overcome

Longino’s attempt to characterize a knowledge‐based consensus solely in terms of social norms a community must satisfy, and that the difficulties with her account lurk any social epistemology that would try to answer the question solely in terms of such norms as well.

Conclusion

In this dissertation, I have set the goal of developing a normative social epistemology that uses ‘knowledge’ as a success term, and specifies conditions our theories and beliefs need to meet in order to qualify as knowledge. In the previous chapters, I have argued that such a normative assessment of our theories and beliefs must be done at the level of the epistemic community rather than the individual agent, and that there are indeed epistemic standards by which we can assess the normative status of our theories and beliefs, which are independent of the our social interests and values.

In this chapter, I proposed such a theory. I focused on the question of when a consensus is knowledge based. I have argued that in order for a consensus to be knowledge based, we must eliminate its being accidental in some way or formed due to a non‐epistemic factors as best explanations of the consensus. I have identified three conditions under which this is likely to be the case: apparent consilience of different lines of evidence, social diversity, and meta‐agreement. I have further argued that the

IBE approach and my suggested conditions in particular overcome difficulties with attempting to identify social‐epistemic norms a community must follow to form a knowledge‐based consensus.

‐ 145 ‐

Having spelled out the details of my theory, one may wonder how it applies to specific instances of consensus. In the next chapter, I will examine a concrete example – the consensus over the safety of the drug Bendectin that emerged in the late 1980s and early 1990s in the U.S. I will show that under my theory, this consensus was not knowledge‐based, and I will argue that this has major implications to the general practice of deference to consensus that is common in public disputes over scientific facts.

‐ 146 ‐

Chapter 4

Was the Consensus on Bendectin Knowledge Based?

Introduction

In the previous chapter, I presented my theory of knowledge‐based consensus. I argued that a consensus is knowledge based when knowledge is the best explanation of the consensus. I identified three conditions under which knowledge is the best explanation of a consensus: apparent consilience of different lines of evidence, social diversity within the consensus community, and meta‐agreement, which amounts to shared acceptance of the same evidential standards and essential background assumptions.

In this chapter, I discuss a case study involving a scientific consensus. In the

1980s and early 1990s, there was a series of mass tort trials in U.S. Federal Courts involving birth defects allegedly caused by a drug called Bendectin, which was manufactured by Merrell. At that time, the scientific community was divided on the question of whether Bendectin could cause birth defects in human babies. Toward the

1990s, however, a scientific consensus emerged that it could not. This consensus was used inter alia as a resource for deciding the case in courts, under the assumption that a scientific consensus amounts to knowledge. The application of my theory to this case, however, reveals that the consensus does not satisfy the conditions for being knowledge based. I therefore argue that it could not legitimately serve as a resource in arbitrating the dispute in court. I further advise general caution in deferring to a scientific consensus in order to decide policy‐related disputes.

This chapter consists of three sections. In section 1, I provide the essential historical and legal background needed for understanding the Bendectin controversy. In

‐ 147 ‐

section 2, I discuss the Bendectin case study and argue that the consensus in this case was not knowledge based. In section 3, I discuss the meaning and of this example and the negative implications of the practice of deferring to a consensus in order to resolve disputes.

1. The Bendectin Trials – General Background

Several sources give thorough and detailed accounts of the history of the Bendectin litigation (Green 1996; Sanders 1998; Edmond & Mercer 2000). My aim in this chapter is not to provide an alternative historical account, but rather to provide an alternative analysis of the role and epistemic significance of the scientific consensus that emerged in this case. In this section I therefore provide only the essential background that is relevant to the issue at hand, rather than a comprehensive historical account.

The Bendectin trials were mass tort trials, namely civil actions involving numerous plaintiffs against one or a few corporate defendants. In the U.S. federal legal system, where the Bendectin trials were conducted, there are three instances. The lower instance consists of 94 district courts, which are trial courts. A higher instance consists of 13 circuit courts, which are appeal courts. Their decisions set binding precedents for district courts in their circuit, but not for other circuits. The highest instance is the U.S. Supreme Court, which is an appeal court, the decisions of which set binding precedents to all lower courts.

In torts cases, in establishing a legal case for compensation, a plaintiff must show that the defendant had a legal duty to prevent harm, the defendant violated it, the plaintiff suffered a legally compensable injury, and the defendant’s action was the factual and legal cause of the injury (Cranor 2005, 143‐144). Toxic torts, i.e. torts cases

‐ 148 ‐

that deal with injuries from allegedly harmful substances, pose unique epistemic problems in proving causation. The causal mechanism behind the injury is often highly complex and poorly understood. There is often long latency period between the exposure and the injury. Most injuries are the result of a long term exposure to an agent at low doses (Parascandola 1997, 147). The subject is often exposed to more that one chemical agent, sometimes produced by different parties, and it is difficult to determine which agent or combination of agents is responsible for the injury. Toxic agents do not always cause harm. They cause harm only at a certain proportion of instances. Some conditions are not uniquely associated with one chemical agent, and sometime occur without exposure to known agents (Simon 1992, 35).

Generally speaking, prior to 1993, the legal test for admissibility of scientific evidence or expert testimony was that it was “generally accepted in the relevant scientific community,” (a criterion established in the Frye [1923] decision) and helpful to juries in their fact‐finding mission. Judges tended to permit experts to testify and let cross‐examination during trial determine whose experts the jury believed. Cross‐ examination in front of a jury was the main screen for expert testimony. In the

Bendectin litigation, starting from the mid 1980s, however, judges started making admissibility tests stricter – and in the Daubert (1993) decision, which was one of the

Bendectin cases, the Supreme Court set new standards for admitting scientific evidence.

In Daubert, the Supreme Court determined that judges should act as ‘gatekeepers’ and admit only ‘reliable’ evidence, which it took to be ‘good science’ that was produced by the application of ‘the scientific method’. They should also admit only evidence that is

‘relevant’, i.e. fits the particular case. According to Daubert, judges should keep ‘bad

‐ 149 ‐

science’ out, as it is unreliable. Daubert recognized four indicators of good science: falsifiability, publication in peer‐reviewed journals, known chance of error, and general acceptance within the scientific community.84

De facto, Daubert institutionalized the stricter admissibility tests and procedures that developed in the course of the Bendectin litigation. Today, a “Daubert” review of experts is a regular procedure in trials of this sort, and it occurs early in the timeline of events leading to a trial. After the initial complaints and answers have been filed, and after discovery, during pretrial hearings, a judge reviews whether the experts will be permitted to testify. Experts for both sides would typically submit reports that provided the basis of their testimony, which would be subject to review by the opponents and by the court. Often one side or the other will file a Daubert motion to have the others’ litigants excluded from testifying. If an expert critical to a litigant’s case is not admitted, the litigant (typically the plaintiff) may be unable to establish needed factual premises, in which case the judge can dismiss the attempted legal action by means of a summary judgment because there is no material issue of fact for the jury to decide. Thus, a preliminary review of whether one’s experts may testify can result in dismissal of the case without a trial. (Cranor 2005, 143‐145).

The Bendectin trials took place from the early 1980s until the mid 1990s While in the early 1980s district courts tended to rule for the plaintiffs, from the mid 1980s they started to rule against the plaintiffs on account of not showing causation. Summary judgements became more common, and circuit courts tended to reverse previous district courts’ jury decision in favour of the plaintiffs. While this was the main trend,

84 For a critical discussion of the Daubert decision and its flawed understanding of what it took to be the relevant issues in epistemology and philosophy of science see Haack (2003; 2004; 2005).

‐ 150 ‐

because courts in each circuit can rule independently of one another, there were also exceptions. The last Bendectin case was the Oxendine (1996) decision, in which the D.C.

Superior Court (an appeal court) reversed a previous decision and ruled in favour of

Merrell.85

The change in courts’ views from accepting the plaintiffs’ claims in the early cases to rejecting them later on was justified inter alia by an emerging scientific consensus that epidemiological studies in humans are the most reliable indicator for detecting causation between Bendectin and birth defects, and that such studies do not indicate such causation. In this case, the consensus that emerged in the scientific community was used to settle a scientific dispute in court. As I have argued in the last chapter, however the mere existence of consensus, even among scientists, is not necessarily indicative of shared knowledge. Thus, the question arises regarding whether the consensus in this case was in fact knowledge based. Assuming that deference to a scientific consensus to settle a scientific dispute in court is justified only when the consensus is knowledge based, another way to put this question is to ask whether the deference to the consensus was justified in this case. In the next section I answer this question in the negative.

2. Was the Consensus on Bendectin Knowledge Based?

According to Huber’s polemical account (1991, Ch. 7), before the scientific consensus emerged, scientists studying Bendectin and testifying to their findings in courts were doing legitimate science. However, after the consensus emerged, scientists still researching Bendectin or testifying in court about its alleged harmful effects were

85 See a table of the Bendectin cases and their outcomes in Edmond & Mercer (2000) at 304‐5.

‐ 151 ‐

motivated by greed and no longer doing legitimate science, but junk science. The problem with this account is that the only reasonable way to interpret the term ‘junk science’ for Huber is ‘science no longer deemed acceptable by the majority of the scientific community’. Unless we have independent reasons to think the majority must be right, Huber’s claim becomes circular: The scientists’ testimony should not have been accepted because it was junk science, and it was junk science because it was no longer acceptable. With respect to greed, as I will show, it cannot be categorically stated that all remaining concerns about the safety of Bendectin were motivated by greed. We should also keep in mind that while some scientists may have been motivated by greed, Merrell had its own financial interests and resources to promote research that pointed to the opposite conclusion and suppress research that did not.

Green (1996) and Sanders (1998) both offer non‐polemical and more nuanced accounts. Similarly, however, they both argue that the deference to the emerging scientific consensus by the courts in order to settle the legal controversy was justified, without addressing the question of whether the consensus was knowledge based or not.

Green praises the legal system for eventually ruling correctly relying on the accumulating reliable epidemiological evidence that indicated lack of causation (Green

1996, 21‐2, 314‐6).

Edmond and Mercer (2000) dispute these claims, and provide a symmetrical analysis of the events. They argue that the question of whether epidemiological evidence is the most reliable for indicating causation and how it should be amalgamated with other types of evidence was one of the questions courts were required to decide.

Thus, they argue, Green retroactively legitimates the court decisions by the standards

‐ 152 ‐

they eventually reached. Furthermore, Edmond and Mercer argue that courts played a significant instrumental role in bringing about the scientific consensus about the supremacy of epidemiological standards, on which they later relied as a normative epistemic source (Edmond & Mercer 2000, 272‐3).

According to Edmond and Mercer, Sanders similarly offers a Whiggish reconstruction of the events, assuming that the final conclusion lay dormant in the evidence, waiting to be discovered by the courts. They criticize both the courts for resting on an inadequate ideal of science as a value‐free enterprise that produces knowledge that is independent of social influences and interests, in particular by the courts themselves, and Sanders’ account of the trials as a ‘sociology of error’ that assumes the correctness of the final outcome reached and works backwards to explain decisions leading to it in terms of rationality and decisions that stood in the way in terms of biases and social interests. As an alternative, they offer a symmetrical reconstruction of the events:

…care should be made to attempt to document and articulate some of the key processes through which the primacy of epidemiology as an integral part of the closure of the Bendectin litigation emerged. Rather than assume that this conclusion somehow lay dormant ready to be discovered once the dust of legal distortions had cleared, we will document how the context of litigation helped constitute the outcome. (Edmond & Mercer 2000, 282; cf. Jasanoff 1995)

Though they offer a more balanced account, in my view, that accords better with the events, Edmond and Mercer’s analysis leaves us dissatisfied. While they are right, as

I will argue, to analyze the events in terms of negotiations and mutual social construction of the outcome by the scientists and the courts, they refrain from passing a normative judgement on this construction. In the social context of mass torts in the U.S., any final outcome reached would have been a social construction negotiated by courts

‐ 153 ‐

and scientists. However, not any outcome reached is equally epistemically justified. In particular, the mere fact that courts played an active role in bringing about the scientific consensus in itself does not necessarily mean that it was not knowledge based.

My analysis aims at taking Edmond and Mercer’s account one step further, namely not only pointing out the contingent circumstances and events that brought about a particular consensus rather than another, but also determine whether it was knowledge based. In what follows, I will present Sanders’ depiction of the events leading to the consensus. As I will show, Sanders’ depiction of the events actually supports Edmond and Mercer’s analysis rather than his own, as he seems to downplay the role of non‐epistemic factors in bringing about the consensus. I will also show how, by my account, the consensus that eventually emerged cannot be said to be knowledge based.

Four main types of evidence were used in the Bendectin trials: structure‐activity

(toxicology), in vitro (experiments on singles cells, organs, tissue samples, etc.), in vivo

(animal studies), and epidemiology (human studies). Each of these types of studies corresponds to a different line of evidence under my framework. Each type of evidence belongs to what Hacking calls ‘a style of scientific reasoning’ – one of seven fundamental ways of reasoning he identifies as constitutive of Western science (Hacking 2002, 159‐

99). Under my account, a style of reasoning constitutes a set of evidential standards the agreement on which is required to satisfy the third condition of meta‐agreement for a consensus being knowledge based. This information is summarized in Table 3.86

86 This table relies on Sanders (1998, 46‐61). See there for full bibliographical references.

‐ 154 ‐

Evidence Style(s) of Benefits Drawbacks Type Reasoning Structural Analogical Model • Cost • Prone to error – very similar Activity • Facilitate explanation and molecules sometimes have understanding very different effects In Vitro Laboratory Style • Cost • Chemicals may behave • Possibility of cross species differently in the living body extrapolation by than in the test tube due to comparing human and their accumulation only in animal samples certain organs or being rapidly metabolized into different substances In Vivo Laboratory Style; • More accurate knowledge • To reduce costs, trials are done Statistical Style and control of the dose on a small population of than human studies animals and so the dose is • Better control of increased, but at very large confounding factors than doses almost every substance human studies becomes harmful • Different physiology in humans and animals – some chemicals are harmful to humans and not animals and vice versa • No single agreed‐upon method of extrapolation to humans Epidemiology Statistical Style • Direct knowledge of • Data is obtained from what causation in humans – people report and depends on fewer problems for their memory. Less reliable extrapolation knowledge of what doses were taken and when • Less control and knowledge of possible confounding factors • Existence of causation at very small rates is hard to identify due to insufficient sample size Table 3: Evidence Used in the Bendectin Trials When talking about the validity of studies, a distinction is drawn between

internal validity, which concerns inference with respect to the population or samples

actually tested, and external validity, which concerns the extrapolation of the results of

the study to the target population. All types of studies are prone to problems of internal

and external validity. Generally speaking, evidence belonging to the laboratory style

tend to be more prone to problems of external validity, as the subjects or samples differ

‐ 155 ‐

in relevant ways from the target population, and the laboratory settings try to create an idealized environment that shields the experiment from influences that may exist in the target population (Sanders 1998, 53‐60).

Another possible problem for in vivo and epidemiological studies that belong to the statistical style is of statistical conclusion validity. One problem is the conditions under which it is legitimate to infer causation from correlation. These conditions vary with circumstances and give rise to disputes, especially in the legal context of tort law that has its own understanding of causation.87 Another problem is the trade off that generally exists in statistical studies between increasing the chances of false positives and decreasing the chance of false negatives and vice versa (Sanders 1998, 50‐3).

The latter issue is particularly important in this case study. In vivo studies of a substance on several types of animals are very effective in preventing birth defects.

There have been no known new drug‐induced epidemics of birth defects since the implementation of laws requiring animal testing (Sanders 1998, 57). While in vivo studies are very effective in preventing birth defects, they do so at the cost of a high rate of false positives, namely they predict more birth defects in humans than there are.

Sanders claims that this feature makes in vivo studies valuable in the regulatory context that aims at prevention of birth defects, but not so much in the legal context that aims at causally explaining existing ones (1998, 58). However, he seems to overlook Edmond and Mercer’s point that the exact level of tolerance of the legal system to false positives was not pre‐determined but constructed and negotiated during the Bendectin trials.88

87 See Parascandola (1997) and Simon (1992) for discussions of the relations between correlation and causation in tort law. 88 The asymmetry between the two contexts also questions Solomon’s (2001, Ch. 2) use of empirical success as the ultimate normative criterion for science. While in vivo studies have great predictive

‐ 156 ‐

From the mid 1980s epidemiological studies conducted failed to replicate the positive results of the previous studies. Sanders argues that within the field of epidemiology, the question whether the studies showed causation was ‘purely factual’

(1998, 86). He argues that

None of the scientists involved in the Bendectin studies had tied the outcome of the research to an important theoretical issue. In this environment, the scientists were not prepared to invest the intellectual energy necessary to provide nonsubstantive explanations for replication failure. Even central players […] were prepared to interpret failure to replicate as a factual finding, and to revise their opinion accordingly (Sanders 1998, 86).

However, Sanders’ analysis conflates two different issues. The results of the studies were not purely factual, but theory laden. For example, as he mentions, an important question in epidemiological studies was in which stage of the pregnancy the mothers took Bendectin. It was theorized that if it were teratogen, Bendectin would work at the organogenesis stages of their pregnancy – the early stages in which internal organs develop, so women who took it later should be excluded from the population study

(Senders 1998, 53). This example shows that theory is involved in the design and interpretation of the study. The second claim is that epidemiologists, even those who initially claimed that Bendectin was harmful, did not have a strong vested social‐ cognitive interest (Pickering 1982) in a particular theory of causation, which might have motivated them to offer an alternative theoretical interpretation of the failed replications.

Sanders agrees that the controversy over the supremacy of epidemiological studies over other types of evidence was theoretical and not purely factual, but argues:

and manipulative empirical success in the regulatory context, they have poor empirical success in the legal context. Empirical success is relative, then, to the social context, while for Solomon it is supposed to serve as an anchor that saves Social Empiricism from SSK relativism.

‐ 157 ‐

If the in vivo studies had clearly pointed in a different direction from the epidemiology, the controversy would not have been so easily resolved. However, controversy concerning the primacy of epidemiological evidence is largely irrelevant to the Bendectin Controversy. The in vivo and aggregate time trend data support the epidemiological evidence that either Bendectin is not teratogen or is such a weak teratogen that its effects are undetectable. In vitro studies are more ambiguous, but it is generally agreed that this evidence is less probative than in vivo and epidemiological studies (Sanders 1998, 86).

This assessment, however, does not accord with the events as Sanders himself depicts them. The majority of epidemiological studies conducted from 1981 to 1994 concluded with a negative result (no effect observed), and a few were inconclusive

(Sanders 1998, 70). By contrast, nine in vivo studies were conducted from 1981 to

1988; three concluded with a negative result, four with a positive result, and two with inconclusive results (Sanders 1998, 66‐7). In 1985, nine toxicologists wrote a letter to the journal Teratology expressing scepticism about whether the safely of Bendectin had in fact been established (Brown et al, 1985; Sanders 1998, 88). Two additional letters in this issue discuss the safety of Bendectin and there is clear dissent within the community of toxicologists (Brent 1985, Holmes 1985). From a social‐epistemic point of view, the existence of this dissent is significant.

Moreover, as Sanders himself mentions, the in vivo studies practically stopped from the mid 1980s, not because the scientific community was satisfied that the epidemiological evidence was better, but because courts were:

The direct evidence on human effects provided by the epidemiological evidence diminished the value of in vivo studies insofar as they are designed to answer the question of whether a drug causes harm to humans. Increasingly, courts simply refused to rely on or even admit in vivo evidence. Considering the declining demand for in vivo research, it is not surprising that the supply has been limited (Sanders 1998, 68; emphasis added).

‐ 158 ‐

In other words, when dissenting views within the toxicologist community called for more in vivo research, it was not done because courts stopped relying on it or admitting it.

It is important to bear in mind that the practice of not admitting in vivo research did not emerge in a vacuum, but in a specific social context. The U.S. legal system is a jury system, where the jury – twelve citizens with no legal background – assume the role of the trial fact finder. This practice was established in a period of growing concerns, especially by conservatives, about jurors’ abilities to carry out their role as fact finders in trials involving complex scientific evidence (Foster et al 1993; Huber

1993). Specifically, the problem arose in mass tort trials, such as the Bendectin trials, where the plaintiffs claimed that their exposure to a harmful substance produced by the defendant caused their illness. In such cases, it was argued that jurors tend to sympathize with the plaintiffs, who are typically ordinary working‐class citizens, and be hostile to the defendant, typically a big corporation. Thus, they tend to decide in the plaintiffs’ favour based on ‘junk science’ or irrelevant evidence presented by the plaintiffs’ experts and purporting to show a causal link between the suspected substance and the plaintiffs’ illness (Sanders 1998, Ch. 1 & 2). Such practice by the jury, it was argued, not only averts justice, but also inhibits economic development, as it compels big corporations to spend major funds for defending themselves in legal trials and paying compensation.89 A good explanation of the emergence of the practice of not admitting in vivo research – namely preventing it from reaching a jury, is that it was a response to such worries.

89 For a discussion and critique of these claims see Cranor (2005).

‐ 159 ‐

Sanders claims that ‘in the case of the Bendectin controversy, it seems certain that nonepistemic factors did play a role in the general consensus that formed around

Bendectin’ (1998, 87). He expresses great confidence in the internal mechanisms of group rationality of the scientific community to resolve controversies correctly, and expresses regret that the legal system lacks such mechanisms (1998, 116). It seems that because of this confidence, though he is aware of them, Sanders seriously downplays the importance of external influences on the scientific community and the mutual relations between it and the legal system. In particular, he downplays the role the court played in blocking further in vivo research, which Edmond and Mercer (2000, 276) find pivotal in constructing the decision about Bendectin. Thus, his appeal to the consensus that was formed as a justification for the court decisions is illegitimate.

My theory of consensus supports this conclusion. The scientific consensus that

Bendectin does not cause birth defects did not exhibit apparent consilience of different lines of evidence. Rather, it was achieved by one line of evidence, namely epidemiological studies, taking over the rest. In other words, in this case, various factors led the legal and scientific communities to favour the negative results obtained through epidemiological studies over the inconclusive results that were obtained through in vivo and in vitro studies, and to abandon further in vivo research in spite of dissenting calls to the contrary. Under my account, the consensus that emerged in the mid 1980s and early 1990s was not knowledge‐based. Hence, the mere existence of consensus in this case could not serve as a justification or a normative resource for courts on which to base their decisions.

‐ 160 ‐

3. The Wider Picture – General Lessons from the Bendectin Case Study

What are we to make of the claim that the consensus in the Bendectin case study was not knowledge based? It is important to distinguish between two questions. The first question is whether Bendectin causes birth defects. This is a scientific question about human physiology. The second question is whether the scientific consensus that

Bendectin does not cause birth defects was knowledge based. This is a meta‐scientific question, and it belongs to the realm of social epistemology. In this chapter, I addressed the second question, not the first.

Generally speaking, it is possible that a proposition p is true, and that there is a scientific consensus that p, but the consensus is not knowledge based. Suppose, for example, that p is true, but all the parties to the consensus are biased toward the truth of p, and would probably all believe that p even if p were false. The opposite situation is also possible. Because my theory of knowledge‐based consensus is fallibilistic, it is possible though implausible under my theory that there will be a knowledge‐based consensus over p, when p is in fact false.

As I have stressed in the previous chapter, the statement that a consensus over p is not knowledge based does not mean that p is false. Rather, it means that the mere fact that an agreement exists in an epistemic community does not carry epistemic weight. It does not give us any reason to believe that p over and above the reasons the members of the community have to believe that p. It may still be the case that these reasons alone are compelling and sufficient for justifiably believing that p, or it may not. When we can confidently say that a consensus is knowledge based, this means that we have an independent reason to justifiably believe that p.

‐ 161 ‐

Applying this general discussion to the Bendectin case, my claim that the

Bendectin consensus was not knowledge based should be neither confused with the claim that Bendectin does not cause birth‐defects, nor with the claim that the courts ruled incorrectly, nor with the claim that the evidence for the theory that Bendectin does not cause birth defects was insufficient for justifiably believing so. Rather, it means that the courts could not justify their decisions by appealing to the fact that a scientific consensus on this matter existed. It also means that the scholars that justify the court decisions by appealing to the emergence of the consensus on this matter are wrong in their epistemic analysis of this case study.

Another question that arises is whether animal research on Bendectin was terminated prematurely. Solomon would seem to answer this question in the positive.

Solomon argues that as long as a theory enjoys some empirical success, some research efforts should be allocated to it. In the Bendectin case, so she would seem to argue, since some animal studies showed positive results, animal research was terminated prematurely.

I am not committed to this position, and my analysis does not necessarily imply that it was wrong to terminate further Bendectin research in order to conclusively determine whether it could cause birth defects. There might be overriding social considerations that would justify terminating further research in this case. Recall that in the mid 1980s, when animal research was de facto terminated because courts no longer accepted stand‐alone evidence from animal studies, Bendectin had already been taken out of the market. It no longer posed a potential threat to public health. It might thus be argued that further public resources in order to conclusively establish whether it

‐ 162 ‐

caused birth defects or not just for the sake of the Bendectin trials should not have been allocated. Such an argument, though, cannot be supported by the existence of consensus, since the consensus in this case was not knowledge‐based. It may be supported by non‐epistemic considerations that concern the just distribution of

resources in society. As such, it exceeds the scope of this dissertation. It follows from my analysis, however, that if such further research had been done, it would certainly not have been ‘junk science’ as Huber claims.

In my view, this cause study constitutes a warning sign against misusing scientific consensus in public debates, and against a demand by the public or by decision makers for a scientific unified front on a given matter as a necessary condition

for warranting action. Requiring a scientific consensus conveys a distorted image of

science as a body that always speaks in one voice. Public demand for consensus may also lead to the undesirable consequence of silencing dissenting voices within the

scientific community, and may simply set too high a bar for science to meet.

Such demands also ignore the normal existence of scientific pluralism and dissent, and their positive epistemic role in justifying our theories and discovering the truth, which, as I noted in the previous chapter, has been often highlighted in the philosophy of science. As Dascal argues, rather than being abnormal and detrimental to science, ‘controversies are indispensable for the formation, evolution and evaluation of

(scientific) theories, because it is through them that the essential role of criticism in engendering, improving, and controlling the “wellformedness” and the “empirical content” of scientific theories is performed’ (Dascal 1998, 147).

‐ 163 ‐

Conclusion

In this chapter, I applied my account to the controversy that arose in the 1980s and early 1990s in a series of legal tort trials in the US over the question whether the drug

Bendectin causes birth defects. I argued that while courts deferred to the consensus that emerged in the scientific community that the drug in fact does not cause birth defects, it could not be said to be knowledge based under my theory of knowledge‐ based consensus. I distinguished between the question of whether the consensus in this case was knowledge based from the question of whether Bendectin was harmful as claimed. I argued that a negative answer to the first question does not necessarily imply a negative answer to the second question, but rather disqualifies the deference to the consensus as a legitimate way of resolving the second question in this particular case.

As I have demonstrated, the consensus in this case did not exhibit an apparent consilience of evidence. Rather, due to certain contingent social settings, epidemiological evidence triumphed over all other types of evidence. A general question thus arises about the possible ways in which social values, interests and the like may influence the process of evidential reasoning. In the next chapter I address this question.

‐ 164 ‐

Chapter 5

The Role of Values in Evidential Justification

Introduction

In the last two chapters I presented and illustrated my theory of knowledge‐based consensus. One of the conditions for a consensus being knowledge based is that it should exhibit an apparent consilience of evidence. This means that all the available evidence must seem to support the consensual view, and no seemingly contradictory piece of evidence should be ignored. This condition aims at guarantying that as a whole, the consensus is evidence based, and not skewed by agents’ irrelevant psychological motivations, ideological views, wishful thinking and the like.

But how do we know when an agent’s belief is mostly evidence based rather than skewed by such irrelevant values? We need a descriptive model from the point of view of the theory of evidence of the logical relations between values, evidence, and theory or belief, and the roles values may play in evidential justification. Current writing on this subject focuses almost exclusively on one such model which describes one such role – values fill the gap of underdetermination between theory and evidence. In this chapter I will argue that the underdetermination model only partly captures the relations between values, evidence, and theory or belief. I will identify three additional roles values play in evidential reasoning and show their significance.

In this chapter I will use a broad notion of values. I consider as a value anything that serves as a basis for discriminating between different states of affairs and ranking some of them higher than others with respect to questions of how the personal, social, natural, and cosmic order ought to be (cf. Taylor 1992, 29‐30). Political, ideological and

‐ 165 ‐

moral values naturally fall under this rubric, but also social interests and psychological motivations, as they discriminate between more and less desirable states of affairs in our personal and social lives. Theoretical or cognitive values, such as simplicity, scope, elegance, predictive power and the like fall under this rubric of values as well, as they rank theories that have certain properties as higher than others. Similarly, I will use a broad notion of evidence as anything that serves as a basis for discriminating between different states of affairs and ranking some of them higher than others with respect to questions of how the world is.

This chapter contains seven sections. In section 1, I review the underdetermination model and its limitations. In section 2, I argue that social values affect the trust we extend to the testimony of others, and that this epistemic role is not another aspect of filling the underdetermination gap. In section 3, I review research in experimental psychology on the influence of values on people’s evidence assessment, which will serve as a springboard for the discussion in the next sections. In section 4, I argue that social values lower and raise evidential thresholds, and in section 5, I argue that they affect the relative weighing of discordant evidence. In section 6, I demonstrate the importance of my model using existing case studies in philosophy of science. In section 7, I suggest that my framework may have major implications for existing theories of the nature of knowledge and epistemic justification.

1. The Underdetermination Gap­Filling Model and Its Limitations

According to the underdetermination thesis, any body of evidence can logically accommodate more than one theory and perhaps infinitely many. Some philosophers of science draw on the underdetermination thesis to argue that values ‘fill the gap’

‐ 166 ‐

between theory and evidence. According to the so‐called gap‐argument, because evidence alone does not determine theory, scientists must use values to choose which theory to accept. They choose the theories that are most consonant with the values they cherish. For example, Longino states:

I take the general lesson of underdetermination to be that any empirical reasoning takes place against a background of assumptions that are neither self‐ evident nor logically true. Such assumptions, or auxiliary hypotheses, are the vehicles by which social values can enter into the scientific judgement. If, in principle, there is no way to mechanically eliminate background assumptions, then there is no way to mechanically eliminate social values and interests from such judgement (Longino 2004, 132‐3).

The claim that values can fill the logical gap of underdetermination of theory by evidence is hardly disputed. It is widely acknowledged in contemporary history and philosophy of science that values, social and other, can and do in fact play a role the process theory choice in the context of justification. A debate exists, however, about the kind of values that legitimately fill the gap of underdetermination, and the extent to which social values influence theory choice by filling the gap in scientific practice.

As for the question of legitimate values, some argue that only cognitive values, such as a theoretical simplicity, elegance or explanatory scope, have a legitimate role to play in theory choice. These values are considered benign and internal to science

(McMullin 1983; Laudan 2004). Others argue that social values, such as political and ideological values, have a legitimate role to play as well. For example, Kourany (2008) argues that our commitment to the social value of racial equality serves as a legitimate reason, all other things being equal, to prefer scientific theories that explain differences in intelligence scores between blacks and whites by appealing to social and cultural factors rather than biological factors.

‐ 167 ‐

The claim that social values have a legitimate role in theory choice is controversial, and has been subject to criticism. Critics argue that it does not follow from the fact that social values play a role in theory choice that they ought to. Social values, so this objection goes, reflect our vision of the desired social order – how the world ought to be, while theories aim at describing the world as it is. Hence, social values are external to the aims of science and constitute biases that need to be eliminated. They cannot serve as legitimate reasons to prefer one theory over another

(Intemann 2005).

Such criticism is common, but rests on problematic assumptions. It presupposes that a sharp, principled, and meaningful distinction between cognitive and social values can be maintained. Proponents of the legitimate role of social values in science question this distinction. Longino (2002, 77‐96) argues that the distinction between social and cognitive values rests on a false dichotomy between the rational and the social as two mutually exclusive domains. In addition, cognitive values are social in the sense they the shared communal values of scientists with a similar training (Machamer and Douglas

1999, 50‐1). Cognitive values may also align themselves in some cases with social and political values, thus in such cases, the choice between competing cognitive values, such as novelty versus external consistency, becomes a choice between competing social values as well (Longino 1995; Machamer and Douglas 1999, 50‐1). Appealing to the aims of science seems unlikely to help maintain the distinction as well. Solomon (2001,

51‐63) argues, from a consequentialist perspective, that cognitive values do not promote the cognitive aims of science, primarily empirical success, more than social

‐ 168 ‐

values. The aims of science themselves are never purely cognitive but also social

(Kitcher 2001, 65).

If social values cannot be sharply distinguished from cognitive values or this distinction is insignificant with respect to attaining the aims of science, and if, as the critics acknowledge, the influence of some values, namely cognitive values, on theory choice is necessary and benign, then it is possible for social values to play a legitimate role in theory choice. I take the arguments to this effect to be largely persuasive, and for the remainder of this chapter I will assume that social values cannot be categorically excluded from the set of legitimate values influencing theory choice. This is not to say that the effect of any social value is legitimate in any context (cf. Douglas 2000, 560;

Wilholt 2009, 96‐7).

While critics of the gap argument do not manage to show, in my view, that social values have no legitimate role to play in theory choice, there is a more challenging criticism of it, which concerns the extent to which social values do and can play a role in evidential reasoning. Norton writes that the gap argument rests on an ‘improvised and oversimplified account of the nature of inductive inference’ (2008, 19). He argues that under most commonly used models of confirmation, when two theories can accommodate the same evidence, they do not enjoy the same inductive support by the evidence, i.e. the same evidence does not equally confirm the two theories and they do not enjoy the same warrant. In many cases, the evidence inductively favours one theory over another (2008, 26‐33). For Norton, this means that the role of social values in filling the logical gap of underdetermination is restricted to relatively rare cases in which theories enjoy a similar inductive support (2008, 20). Haack adds that even when

‐ 169 ‐

two empirically equivalent theories enjoy the same warrant, scientists need not appeal to social values to choose between them – they can simply withhold judgment (1998,

110‐11).

While Norton and Haack may be right in pointing out a deficiency in the gap argument, it does not follow from their criticism that the role of values in evidential reasoning is as limited as they suggest. First, different inductive inference methods sometimes favour different theories. Thus, when the choice between theories is politically charged, the choice of inference method may become political as well and values may affect it, even if the inference method itself does not have inherent social or political content. For example, an analysis of statistical data using the method known as

‘vote counting’ shows that there is no relation between monetary resource inputs and students’ educational achievements, while an analysis of the same data based on combined significance tests shows a significant correlation (Hedges et al., 1994). While the choice between these two statistical methods is arguably not political in itself, in this case it may become political and may be influenced by social values because the question whether the state should allocate more money to schools is politically charged.

Another and more significant reason this criticism of the gap argument does not establish that the role of values in evidential reasoning is limited is that the gap model does not exhaust the role of values in evidential justification. Values are involved in evidential justifications in other ways, which have been alluded to by historians and philosophers of science but not systematically discussed. The gap argument is incompatible with Kuhn’s view of the role of values in science. While the underdetermination model is sometimes attributed to Kuhn (1977) who argues that

‐ 170 ‐

shared values of the scientific community direct the community’s theory choice, a careful reading shows that Kuhn in fact does not talk about values as filling the underdetermination gap in the way discussed above. Under the gap‐filling model, when there are two rival theories that can equally accommodate the same evidence, scientists appeal to values to decide between them. According to Kuhn, however, during a transition period between paradigms there is a large but never complete overlap between the problems that can be solved by the old and the new paradigm. In crisis periods, scientists are faced with two rival theories none of which can accommodate all the evidence (Kuhn 1970, 85). Moreover, scientists do not choose the new theory because it can better accommodate the existing evidence, but because they have faith in its potential, still unrealized, to eventually do so:

The man who embraces a new paradigm at an early stage must do so in defiance of the evidence provided by problem‐solving. He must, that is, have faith that the new paradigm will succeed with the many large problems that confront it […] A decision of that kind can only be made on faith. (Kuhn 1970, 157‐8).

Moreover, according to Kuhn, in periods of crisis scientists re­rank the evidence that they have with respect to its importance. Minor problems that were once thought to be solved or capable of being eventually explained suddenly become pressing counter evidence against the old theory (1970, 66‐75). The way values play a part in this re‐ ranking process has not been adequacy characterized.

The gap‐filling model, then, fails to seriously take into account the role of inductive confirmation in theory choice, and is incompatible with Kuhn’s influential view regarding the role of values in theory choice. In the next sections I will suggest a new model of the logical relations between values and evidence that can rise to these challenges.

‐ 171 ‐

2. Social Values Affect Trust in Testimony

Rolin (2004) argues that social epistemologists have focused on the way social values fill the underdetermination gap to the neglect of other epistemic influences thereof. She identifies another role social values play, which is affecting people’s trust in each other’s testimonies. As you recall, in Chapter 1, I defended and developed Hardwig’s (1985) argument according to which individuals qua individuals, scientists included, do not possess direct evidence for most of their beliefs. Only an epistemic community, whose members irreducibly trust in each other’s testimonies, jointly possesses such evidence.

Drawing on Hardwig’s argument, Rolin points out that if science is to be successful in achieving knowledge, scientists’ trust in each other’s testimonies should be based on justified confidence in their colleagues’ cognitive authority in their respective domains of expertise. She argues that social values may affect scientists’ assessment of their colleagues’ credibility and trustworthiness. For example, historical case studies show how gender bias has caused male scientists to unjustifiably disregard or dismiss female scientists’ testimonies and overrate testimony by other males (Rolin 2004, 881‐4).

In this section, I will expand this claim and argue that the effect of social values on trust in testimony is indeed distinct from their role in filling the underdetermination gap. I will argue that the process of forming a belief based on a person’s testimony is not, at least in many cases, a process that involves any theory choice. Hence, if values affect this process, their effect cannot be another aspect of their role in filling the underdetermination gap.

Before we evaluate this claim, let us first examine what is involved in trusting another’s testimony. According to Baier’s influential account, trust is reliance on

‐ 172 ‐

another’s good will toward one. It is distinguishable from other forms of reliance on another’s behaviour, which are not directed at that person’s good will. When one depends on another’s good will, one is necessarily vulnerable to the limits of that good will. Under this account, trust is one’s accepted vulnerability to another’s possible but unexpected ill will (Baier 1986, 235). Baier’s account needs to be supplemented with other components. First, in some cases it seems that good will is not necessary, only lack of ill will. Second, there are instances in which one extends or denies trust not as reliance on another’s good will but as reliance on his competence. For example, I may not trust a colleague with some expensive and fragile equipment, not because I suspect his good will, but because I suspect his ability to handle the equipment with enough care.

Respectively, epistemic trust is often expressed in terms of reliance on one’s sincerity and competence (e.g. Fricker 1995, 398). When a person is sincere with me, he shows epistemic good will (or lack of ill will) toward me. He does not act to inhibit my achieving epistemic aims such as gaining truth or knowledge. He might, however, still think that it would be best for me for non‐epistemic reasons that I did not achieve these aims. For example, he might think that it would be best for my psychological well being that I did not know that my wife was cheating on me. The notion of epistemic trust, however, may still be cashed out in terms of sincerity and competence.

Returning to the claim, according to which social values affect the trust scientists extend to each other’s testimonies, one may argue that this is just another way in which they fill the underdetermination gap. This seems to be Longion’s position. Longino argues that the underdetermination thesis entails that certain social norms are

‐ 173 ‐

necessary for the successful generation of knowledge (2002, 124‐8). One of these norms is ‘tempered equality of intellectual authority’, which states that ‘the social position or economic power of an individual or group in a community ought not determine who or what perspectives are taken seriously in that community’ (2002, 131). Rather, only relevant expertise ought to be taken into consideration. Tempered equality aims at achieving a diversity of perspectives which is necessary for epistemically effective critical discourse (Longino 2002, 131). The idea seems to be preventing situations in which there are several theories that can accommodate the same evidence, but some of them are not seriously considered because their proponents are socially disempowered.

For example, when both cultural and biological explanations are suggested to explain differences in intelligence tests scores between men and women, but only biological theories are considered because female proponents of the cultural theories are dismissed due to gender bias. Under this view, the justification for tempered equality is preventing social values from negatively affecting theory choice. This view regards the influence of values on trustworthiness assessments as merely another way in which the fill the underdetermination gap.90

In my view, however, this is not the main justification for tempered equality.

Tempered equality is desirable even if it does not bring about diversity of perspectives.

Tempered equality is a special manifestation of a more fundamental epistemic principle according to which attaining knowledge depends on reliable sources. If scientists overrate or underrate the reliability of their instruments they will get unreliable data. In the same manner, it is desirable that scientists accurately assess the reliability of their

90 Rolin (2004, 882) makes a reference to Longino’s justification, but does not address the fact that for Longino, this justification is one of the lessons of the underdetermination argument.

‐ 174 ‐

peers. To know under what conditions and to what extent they are entitled to trust their peers, we want scientists to accurately assess their trustworthiness. Since, as I have argued, social values may skew such assessments, tempered equality is desirable.91

One may insist that the effect of values on trustworthiness assessments is another way in which they fill the underdetermination gap, because when a hearer decides whether to trust a speaker’s testimony, she in effect constructs and entertains in her mind different theories about the speaker’s trustworthiness that fit the evidence she has at her disposal. She then chooses one of these theories, and comes to believe or disbelieve the speaker’s testimony based on that theory. This objection hinges on the view that testimonial beliefs are inferential and trust is a propositional attitude. In what follows I will explore the arguments for and against this view. I will argue that the arguments for this view are not conclusive. Hence, inasmuch as testimonial beliefs are not inferential and trust is not a propositional attitude, the role values play in influencing trust in testimony is distinct from their role of filling the underdetermination gap.

Some philosophers claim that testimonial beliefs are inferential. Thagard (2006) and Lipton (2007) suggest models in which a person’s trust in a speaker’s testimony is

91 It is important to stress that notions such as tempered equality do not aim at eliminating social values from trustworthiness assessments. Rather, they wish to define the proper way they ought to affect such assessments. At first blush, so‐called logical fallacies such as ad hominem, which connect the validity and soundness of an argument with the identity of the person making it, are underpinned by a similar principle. However, from the perspective of social epistemology, it is not clear that they are logical fallacies. They presuppose an autonomous individual knower who has direct access to all the evidence needed to assess an argument abstracting away from the social context. As I have argued in Chapter 1, normal adult individuals do not have such evidence. As Fuller argues, so‐called logical fallacies may be seen as useful heuristics for deciding whom to trust. For example, ad hominem, ad verecundiam, and ad populum state that the kind of person making a claim, the authority backing a claim, and the popularity of a claim are relevant to determining the validity of the claim, respectively (Fuller 2007, 115). The last insight motivated Chapter 3 of this dissertation, which sought an answer to the question of when a consensus is knowledge‐based.

‐ 175 ‐

based on conscious or unconscious inference to the best explanation of why the speaker is saying what he is saying. Lipton gives the following example:

A man rang my doorbell and claimed that my rain gutters are loose. Should I believe him? They look fine to me, I know that he hasn’t been up on the roof to inspect them properly, and I am further discouraged by the fact that he wants me to pay him today to fix them tomorrow. So I infer that the reason why he said what he did is not because he knows that my gutters do need work (though perhaps they do), but because he is hoping to make a fast buck. This unkind explanation does not entail that what he said is true, so I don’t believe him (Lipton 2007, 244).

According to the inference to the best explanation model, one usually comes up with several possible explanations, and infers the truth of the explanation he deems the best. In the example above, we can think of a choice between two empirically adequate theories: T1, according to which the man is telling the truth, and T2 according to which he isn’t and merely hoping to make a fast buck. Because the hearer deems T2 a better explanation, he infers its truth and decides not to trust the man. As we have seen in

Section 0 1, social factors, such as social stereotypes, may affect the choice between T1 and T2. If this is the case, so the objection goes, the role of values in affecting trust in testimony boils down to filling the underdetermination gap.

How strong is this objection? Let us first notice that Thagard and Lipton do not argue that inference to the best explanation is always involved in trustworthiness assessments. They suggest ‘default‐trigger’ models, in which in most circumstances a hearer accepts what she is told without engaging with any conscious evaluation or inference, but in some cases there are some triggers, such as reasons for suspicion, which cause the hearer to switch to an evaluative mode (Lipton 2007, 240‐1; Thagard

2006, 297‐8). It is consistent with their models that a hearer’s social values influence what factors will serve as triggers for the hearer. For example, the mere fact that a

‐ 176 ‐

speaker is black may serve as a trigger to engage in such inference for a racist white hearer, whereas it would not serve as a trigger for other people.

Second, as Lipton notes, testimonial inference is more modest than scientific inference, and more akin to an intuitive judgement than a careful evaluation of the evidence:

One difference between mundane cases of testimony and some of the more obviously inferential activities of the scientist or the detective is that scientists and detectives sometimes have to do considerable work to come up with the explanations they will go on to infer, whereas in mundane testimony the belief we come to acquire is given to us on a plate, since it is simply the content of the testimony itself (Lipton 1998, 25).

Third, the view that testimonial beliefs, even when they are subject to a reflective process, are inferential is controversial. Audi (1997) argues that testimonial beliefs need not be inferential, even in cases that go beyond default acceptance. He describes a hypothetical case in which a woman of a plane tells him a story about a philosopher who lost his temper at a conference. At first, he suspends judgement, but as the story advances he starts believing her, as she and the story seem more credible to him. At no place in the conversation, so he argues from introspection into the mental process that he has undergone, has he engaged in any inference. Rather, he has simply gradually come to believe her. He has realized his disposition to believe her, and has come to trust her.

Audi’s account does not deny that the brain may be engaged in subconscious information processing, but not in inference. Inference entails belief formation, and at not time, so he argues, has he formed a belief that the speaker is trustworthy (Audi

1997, 407‐9). Underlying Audi’s model is a conception of trust as a non‐propositional attitude, namely a stance. Under this conception, trusting somebody and believing that

‐ 177 ‐

he is trustworthy are two distinct mental states. This conception enjoys popularity among some philosophers. Holton (1994, 67), for example, says the following:

When you trust someone to do something, you rely on them to do it, and you regard that reliance in a certain way: you have a readiness to feel betrayal should it be disappointed, and gratitude should it be upheld. In short, you take a stance of trust towards the person on whom you rely. It is the stance that makes the difference between reliance and trust.

Holton’s argument is based inter alia on his observation that unlike belief, trust seems to be, at least in some cases, in our voluntary control. For example, in a popular exercise in drama class, when you let yourself fall back, there is a moment, so Holton claims, at which you voluntarily decide whether to trust your partner to catch you (1994, 63).

Similarly, Lahno (2001) argues that trust is an emotional attitude, which is a species of stance, rather than a species of belief. He argues that the emotional nature of trust distinguishes it from mere reliance. Lahno argues that trust is a ‘participant attitude’ (Strawson 1974), in which a person regards herself and another person as being involved in interaction. A participant attitude toward another person is characterized by the disposition to have certain emotions toward that person (Lahno

2001, 181). Emotions affect the way a person sees the world. For example, a person in love tends to see the world through pink glasses. A person in love or in another emotionally exhilarating state has the readiness to interpret everything positively, and a person in a gloomy state has the opposite readiness. Having such readiness is not the same as believing that the world is a good or a bad place. It is a stance that we take toward the world, rather than a belief about the world. As an emotional stance, trust has a similar effect. For example, a person whose good friend has been accused of committing a crime may trust her testimony in spite of the incriminating evidence

‐ 178 ‐

because of their friendship and his affection toward her (Lahno 2001, 175‐8). This emotional character of trust seems to be missing altogether from Lipton’s and Thagard’s inferential models discussed above, which take trust to be a propositional attitude.

It is not clear if such philosophical arguments can conclusively settle the debate about the nature of trust, i.e. whether it is a propositional attitude or a stance. In fact, in my view, there is no need to pose the question of the nature of trust in ‘either‐or’ terms.

Some sociologists (Lewis & Weigert 1985) argue that that only a multifaced conceptual analysis of trust is able to account for its explanatory role in sociological theories of the different roles it plays in our social life – trust in testimony, trust in personal relations, trust in institutions, etc.92 They argue that trust ‘has distinct cognitive, emotional, and behavioural dimensions which are merged into a unitary social experience’ (1985, 969).

In particular, they argue that experimental psychology analysis of trust is too narrow.

Experimental psychologists who study trust in various ‘prisoner dilemma’ games look only at the behavioural aspects of trust and its manifestation as rational expectation, whereas in real social life, it has other dimensions, such as emotional, and people often do not have sufficient information for rational prediction (1985, 974‐8).

According to my argument in this section, at least in many cases, the effect of values on the trust we extend to the testimony of others is not another manifestation of their role as filling the logical gap of underdetermination of theory by evidence. For the purpose of this argument, it is sufficient to acknowledge that trust in testimony is, at least in some cases and in some part, a stance with an emotional component rather than an inferential propositional attitude. In those cases, where trust is a stance rather than a

92 Lewis & Weigert’s argument about the nature of trust from the explanatory role of trust in sociological theory might help persuade readers who are not impressed by arguments from introspection, such as Audi’s and Holton’s arguments.

‐ 179 ‐

propositional attitude, values cannot be filling the underdetermination gap, because the hearer is not entertaining any theory about the speaker’s trustworthiness. In such cases, trust does not take the form of a theory in the hearer’s mind that states that the speaker deserves his trust – a theory that would constitute a reason for the hearer to form the corresponding testimonial belief. Rather, in those cases, trust takes the form of a readiness to believe or disbelieve the speaker.

As Miranda Fricker argues, however, if we adopt a non‐inferential model of testimonial belief, we are faced with a problem of explaining ‘how the hearer can exercise sensitivity to the balance of reasons for and against acceptance without making an inference’ and how she may ‘critically filter what she is told without active reflection of any kind’ (2007, 69).

To answer this question, Fricker draws on ideas from virtue ethics and virtue epistemology. She proposes a mechanism of a testimonial perceptual capacity she calls testimonial sensibility, which may explain how people take a stance of trust without engaging in inference that involves theory choice. Testimonial sensibility is a dispositional trait to react in certain ways in certain circumstances. When a person’s testimonial sensitivity functions optimally, i.e. when a person correctly assesses the trustworthiness of a speaker, it is a virtue and that person is virtuous (Fricker 2007, 71).

Fricker notes several features of testimonial sensibility that are relevant to my argument, according to which the effect of values on trust in testimony is not another aspect of filling the underdetermination gap. First, as we have seen, it is not inferential.

Second, it is not codifiable. That is, neither is it based on the application of a theory or set of rules specifying when to trust a person, nor can it be formulated as such a set of

‐ 180 ‐

rules. Third, it has an emotional component. Testimonial sensitivity entails certain emotions such as sympathy, suspicion, respect or contempt. Such emotions reliably guide the virtuous person regarding whom to trust (Fricker 2007, 75‐80). Fourth, testimonial sensibility is socially­situated and may be trained and cultivated in a person’s lifetime:

The Deliverances of an individual’s sensibility, then, are shaped by a set of background interpretive and motivational attitudes, which are in the first instance passively inherited from the ethical community, but thereafter actively reflected upon and lived out in one or another way by the reflective individual (Fricker 2007, 82).

What is the relation between social values and perceptions and the virtue of testimonial sensibility? Current theories in social psychology stress the use of heuristics in social judgments. Hearers make use of social stereotypes as heuristics to facilitate their judgment of a speaker’s credibility. Fricker argues that using their testimonial sensibility they perceive the speaker as trustworthy to a certain extent. Some of the stereotypes they use are unreliable, for example, that women are generally incapable of abstract thinking (2001, 31‐6), and some are reliable, for example, that second‐hand car salesmen are typically dishonest when it comes to selling cars, and that speakers that avoid eye contact, frequently look askance, and pause self‐consciously as if to make up their story are typically insincere, or so she argues (2007, 41‐2). When speakers hold stereotypes as prejudice, i.e. they do not change them in spite of counter‐evidence, they are not virtuous (2007, 35).93

93 In Chapter 1, I argued that the social environment does not provide individuals qua individuals with sufficient evidence or cues about the trustworthiness of other people and their testimonies that amount to epistemic justification. Therefore, I hold that although an epistemically virtuous person who has a well‐trained testimonial sensibility may be in a better epistemic and ethical position than that of a non‐virtuous person, she is still not justified qua individual in most of her testimonial beliefs.

‐ 181 ‐

Prejudice constitutes an impediment to the successful generation and dissemination of knowledge. As Fricker (2007, 43) notes: ‘prejudice presents an obstacle to truth, either directly by causing the hearer to miss out on a particular truth, or indirectly by creating blockages in the circulation of critical ideas’ (2007, 43).

Systematic distortions in the social perception of certain people as more or less trustworthy than they really are typically related to social power. The powerful are considered more trustworthy than they really are and the disempowered are considered less trustworthy than they are. For example, in seventeenth century

England, being a gentleman was a sufficient condition for being considered as generally trustworthy. By contrast, women and lower class men were considered to be generally untrustworthy. This had the practical implication of the exclusion of women from science and the denial of lower class men from getting recognition for their work as laboratory workers, as their claims were considered non‐credible until verified by the testimony of a gentleman (Shapin 1994; Fricker 2007, 119‐20).

Such practices of exclusion also cause the knowledge of the day to support the interests of those in power, portray them in a positive light, and portray those without power in a negative light. For example, the Unites States Supreme Court Decision Plessy v. Ferguson (1896), which stated that racial segregation was constitutional, reflected the biological theories of the time, according to which black were inferior to whites. These theories reflected the socially disempowered position of blacks and their lack of credibility as knowers (Southern 1987, 147). Similarly, Charles Darwin regarded the ability to explain why women have lower intellectual capacities than men as one of the strengths and merits of his theory (Darwin 1871, 326). This view reflects the social

‐ 182 ‐

inequality between men and women in Victorian England. Many other examples are available.94

In this section I have argued that social values affect people’s trust in each other’s testimonies. I have argued that if we grant that testimonial beliefs, at least on some occasions and in some part, are not inferential, and trust is a stance rather than a propositional attitude, then social values play a different role from filling the underdetermination gap. Drawing on Fricker, I have argued that this effect of social values may have detrimental consequence to the quality of the beliefs we hold. In the next section I will argue that in the same way social values affect our stance of trust toward people, they affect our stance of trust toward evidence itself. In sections 5 and

5 6, I will identify two ways in which they may do that.

3. Values, Evidence and Motivated Reasoning

In the previous section, I drew on Fricker’s notion of testimonial sensibility to argue that values play a role in belief formation and theory acceptance that is different from filling the gap of underdetermination of theory by evidence, namely they affect the stance of trust people take toward different speakers’ testimonies.

Sometime, however, the identity of the person or persons who are the source of the testimony is unknown or immaterial to our assessment of the evidence, but this assessment is still similarly influenced by social values. Fricker (2007, 34‐5) briefly discusses such a case, but does not dwell on it. She describes a case in which a panel of blind referees in a scientific journal are prejudiced against a certain scientific method,

94 For racial theories of intelligence in the 19th and 20th century see Gould (1996); for racial biological theories in the 19th and early 20th century see Bowler & Morus (2005) at 415‐437; for gender bias in the biological sciences in the 19th and 20th century see Okruhlik (1998); and for gender bias in science since the Enlightenment see Bowler & Morus (2005) at 487‐510.

‐ 183 ‐

rather than a person. They resist the evidence because of some countervailing motivational investment – they may be conservative and feel loyal to the old methods and not be sensitive enough to the benefits of the new method, or they may be threatened by intellectual innovation. Similarly, they may have a positive prejudice toward other methods.

Such scenarios are common and familiar. As I noted in Chapter 2, much of the work in SSK, which takes the form of detailed case studies, including ethnographies of scientists in the lab, aims at showing that social values and interests are largely responsible for the content of scientists’ beliefs and theories. As I have also noted, this work is controversial, and some philosophers argue that the same case studies can be better analyzed and explained in terms of epistemic notions such as rationality, justification, and truth, rather than social values and interests. Since these SSK studies are controversial, I will not be relying on them to discuss the ways in which social values affect evidential reasoning, as I want my account to avoid the charge of being question begging. Rather, to motivate my discussion in the next two sections I now will present some relevant empirical work in experimental psychology that studies the effect of values on reasoning processes.

The influence of values on evidential reasoning is well documented and studied in experimental psychology. Psychologists have studied the influence of people’s values, preferences and incentives on the ways they form beliefs. The term ‘motivated reasoning’ is used to describe any process of reasoning which is affected by a person’s preference, wish or desire that concerns the outcome of the reasoning process (Kunda

‐ 184 ‐

1990, 480).95 Motivated reasoning is a species of confirmation bias, a concept that refers to people’s tendency to form beliefs in a way that reaffirms their prior beliefs and existing biases (Nickerson 1998; Klayman 1995). Not all cases of confirmation bias, though, are cases of motivated reasoning. For example, when subjects are asked to find the rule governing a series of numbers, they often raise only guesses that confirm their initial hypothesis and do not try to falsify this hypothesis with other guesses. While they exhibit confirmation bias, they do not have any preference toward any outcome, hence this are not cases of motivated reasoning.96

The phenomenon of motivated reasoning is consistent with the discussion in the previous section about the influence of values on people’s assessments of a person’s competence. For example, when faced with the same evidence about a person’s success records in answering trivia questions, subjects, who were asked to participate in a trivia match, tended to evaluate the person’s trivia abilities higher when the person was supposed to be on their team than on the opposite team (Kunda 1990, 487).

Motivated reasoning and evidence assessment are closely linked. People often assess the same evidence differently based on their directional goals. Here are some representative examples, which will motivate the discussion in the next two sections.

When female caffeine consumers read a scientific article claiming that caffeine was risky for women, they were less convinced by it than female non‐caffeine consumers.

Male caffeine and non‐caffeine consumers were similarly convinced by the paper, which controls for the possibility that the results were mediated by prior beliefs rather than

95 The following discussion of motivated reasoning draws heavily on the extensive and influential survey of empirical studies provided in Kunda (1990). 96 The effect of people’s motivation on belief formation does not necessarily entail that people have voluntary control over their belief formation processes. For an overview of the debate on doxastic voluntarism see Vitz (2008).

‐ 185 ‐

by motivational goals. Similar results were found with respect to smokers and non‐ smokers presented with the same evidence about the risks of smoking (Kunda 1990,

489‐90). In another study, sports fans were told that a previously winning team had lost a game. Fans of the team tended to see this as a fluke that did not imply anything for the future. Fans of the opposite team tended to see this as a turning point (Kunda 1990,

488). In a famous study, proponents and opponents of capital punishment received the same mixed evidence about the efficacy of that practice. They ended up further apart and polarized in their views than they had originally been. Similarly, when presented with the same studies, scientists tended to evaluate the studies with the results with which they concurred as more methodologically sound (Klayman 1995, 394‐5).

A point which will be of great significance for the discussion in the next two sections is that although directional goals notably influence subjects’ assessments of the evidence, this influence is also constrained. Subjects motivated to arrive at a particular outcome are not at liberty to conclude whatever they want. Rather, they are constrained by their ability to rationalize their reasoning. Subjects ‘attempt to be rational and to construct a justification of their desired conclusion that would persuade a dispassionate observer. They draw the desired conclusion only if they can muster up the evidence necessary to support it’ (Kunda 1990, 482‐3). Subjects insist on maintaining an illusion of objectivity – they want to appear to themselves and others as following seemingly rational reasoning processes. As Kunda notes, however, this objectivity is illusory because the relevant studies consistently show that

‐ 186 ‐

people do not realize that the process is biased by their goals, that they are accessing only a subset of their relevant knowledge, that they would probably access different beliefs and rules in the presence of different directional goals, and that they might even be capable of justifying opposite conclusions on different occasions (Kunda 1990, 483).

How is it possible for subjects to treat the same evidence differently? How can a smoker and a non‐smoker, for example, treat the same evidence about the dangers of smoking differently? Psychologists have answered to this question by identifying a number of cognitive methods such as selective accessing of different subsets of

memories on different occasions and choosing those reasoning methods that are likely to lead to the desired conclusion (Kunda 1990, 486‐9). These are, however, only partial answers the question. It is still puzzling from the point of view of epistemology and the theory of evidence how such differential treatment of the same evidence is possible.

One explanation, which I discussed in section 1, is that there exists a logical gap between theory and evidence, which allows values to enter and affect subjects’ choice between rival hypotheses. As I have stressed, this is only part of the answer, which cannot account for the vast variety of the cases in which values are involved in evidential reasoning. I will identify additional ways which enable people to treat the same evidence differently. In the next section, I will argue that values may raise and lower the threshold of evidence people require for belief or acceptance. This explanation will apply to the differential treatment of evidence in cases such as the studies about the caffeine consumers, smokers, and sport fans. In section 5, I will argue that values affect the relative weighing of discordant evidence. This explanation will apply to cases such as the study about the mixed evidence on capital punishment and

‐ 187 ‐

the study in which scientists assessed the same evidence differently based on their preferred outcome.

4. Social Values Lower and Raise Evidential Thresholds

In this section I argue that values lower and raise evidential thresholds. That is to say, they affect the strictness of threshold tests that evidence is required to meet in order to support a given hypothesis. Drawing on various examples from the sciences I will explain how this is done. I will argue that this is influence is not another aspect of values filling the underdetermination gap, and that this influence may occur both at the level of the individual person and at the level of the epistemic community as a whole.

In order to understand how values affect the strictness of evidential threshold tests, let us examine how physicists discriminate between signal and noise. In the physical sciences, raw data are never pure. There is always some level of noise, which needs to be removed or reduced. For example, Figure 1 is of a high energy event in a bubble chamber. The left hand side is a photograph of the event – the raw data, and the right hand side is a drawing of the event alone, with the noise removed.

Figure 1: Noise and signal in a high energy event (Brown 1994, 128)

‐ 188 ‐

As Brown (1994, 129) writes, referring to Figure 1, ‘theories explain what is happening on the right; they never try to cope with the mess on the left’. In general, theories do not explain the data in their entirety. They explain the data after the data have been subject to great deal of analysis, which aims at separating the phenomena of interest from extraneous background noise (Woodward 1989, 397).

Scientists’ search for patterns in the data is obviously theory‐driven. They need to know what they are looking for, and this knowledge comes from theory. But the production of patterns is done not only by actively searching for patterns in the data, but also by getting rid of noise. As Hacking observes, in scientific practice ‘“noise” often means all the events that are not understood by any theory’. The process of eliminating or reducing background noise, by debugging measuring instruments , for instance, is not a theoretical activity (1983, 265). Because when physicists distinguish signal from noise, they need not and usually do not try to come up with theories that will explain away the noise, they need not and usually do not engage in any theory choice. This means that, if values influence the process of distinguishing signal from noise in those cases, this is not another aspect of filling the gap of underdetermination of theory by evidence.

Do values affect the process of distinguishing signal from noise, and how?

Frederick Grinnell (1999) argues that the boundary between fabrication of data and creative insight is not always obvious. Scientists are under constant social pressure to present their research in the strongest and most promising way in order to overcome their peers’ scepticism, get funding, and publish. It is hard to do so when they are not genuinely persuaded themselves. It may be hard to identify the exact point in which

‐ 189 ‐

legitimate self persuasion turns into irresponsible conduct. The difficulty stems inter alia from the fact that the boundary between signal and noise may not be clear‐cut:

In research at the edge of discovery, the difference between data and noise often is not obvious. Discovery at the forefront of knowledge requires learning to recognize something when one doesn’t know beforehand what it looks like. Choosing what counts for data will depend on an investigator's experience and intuition – in short, his/her creative insight (Grinnell 1999, 207).

Scientists usually tie their own success with the success of the theory they support. In their assessment of the evidence for the theory that they support, they are influenced by values such as the social pressure to publish, their desire to get research grants and peer recognition, and their fear of failing to achieve these objectives, and damaging their future career. In cases such as the ones described by Grinnell, scientists who have a vested interest in some theory or are otherwise inclined or motivated toward a certain outcome, will tend to see signal when others see noise, and vice versa.

Because the threshold between signal and noise is not clear‐cut, different scientists who adhere to different values are able to differently treat the same data. As I noted in the previous section, this ability is restricted by their respective ability to rationalize to themselves and to others their differential treatment and maintain an illusion of objectivity.

Why is the distinction between signal and noise not clear cut? Mathematically, any data set can be regarded as a sum of two components: a relatively simple and regular pattern and a certain level of noise.97 This means that any data set can formally be described as the sum of one of infinitely many distinct patterns and a corresponding

97 This is a consequence of the fact that any mathematical function f can be represented as a sum of two functions g and r, such that f(x) = g(x) + r(x), g is a regular function, and r is the difference.

‐ 190 ‐

incidence of noise (McAllister 1997, 219‐220).98 The same data set may therefore lend themselves to more than one partition between signal and noise, and scientists, who are influenced by various social values, may debate this partition. In this sense, the decision scientists eventually make is socially negotiated.

The effect of social values on our assessment of the strictness of threshold tests that evidence is required to meet is not limited to the signal‐noise distinction. The choice of threshold values of statistical significance is another example. Statistical studies use various mathematical metrics to evaluate the evidence. One such metric, which is known as significance level (α) or critical p value, is, roughly speaking, the acceptable threshold value for the probability of wrongly accepting a hypothesis.

Another metric is relative risk (RR), which is the ratio of the probability of a disease occurring in a group exposed to a suspected harmful substance to the probability in a non‐exposed group. Only when the relative risk is high enough can it be said that the exposure to the substance is responsible for the occurrence of the disease.

There are commonly used numerical values for such metrics. Very often, a significance level of five percent is used to decide whether to accept or reject a hypothesis. According to one interpretation, these values represent an objective invariant threshold all evidence must meet to be significant. This interpretation was endorsed in a number of U.S. federal courts’ legal decisions in the Bendectin litigation I

98 Woodard (1989) and Brown (1994, 117‐141) argue for a realist interpretation of phenomena, namely the stable patterns that are reconstructed from the data, as things that objectively exist in the world. McAllister argues that within data sets, there is no way to draw a principled distinctions between patters that correspond to phenomena that objectively exist in the world and patterns that do not, and hence, Woodward and Brown’s realist interpretation of phenomena is wrong. In my argument I point out the difficulties scientists face in distinguishing signal from noise and their susceptibility to social negotiations, while I remain agnostic about the ontological status of phenomena.

‐ 191 ‐

discussed in the previous chapter. In these decisions, courts have categorically ruled that meeting the threshold of five percent significance level and a relative risk of two are required for admitting epidemiological studies about the harmful effect of a substance, and any study that falls below of these threshold values is insufficient for showing causation (Beecher‐Monas 2002, 64).

Epidemiologists regard this rule as mistaken. While in the Bendectin trials, the use of significance testing for establishing causation was in dispute by experts on behalf of the plaintiffs and the defendant, the experts on both sides warned the court against a rigid and dogmatic application of such threshold values. They agreed that it is wrong to categorically accept studies whose significance level is lower than five percent and reject all others (Edmond & Mercer 2000, 296‐8). Epidemiologists Bryant & Reinert

(2001, S32) express similar views: ‘0.05 is an arbitrary number […] If scientists and courts insist on “significance” testing, the actual p value should be reported and considered, not simply whether it falls above or below an arbitrary point’ .

But if this is the case, the question remains: What do numerical values such as five percent stand for? Hacking provides a different answer. He argues that such common standards are ‘a technology of intersubjectivity’. When a statement reports that the significance level of a test is five percent, its primary role is not to say something about this test – its aim is not to say what would happen in 95 the cases if the test were taken 100 times. Rather, its aim is to indicate that a certain protocol has been used and to provide a method for qualitative, intersubjective, inter‐test comparisons

(Hacking 1992, 152).

‐ 192 ‐

Similarly, Wilholt regards such threshold values as conventional standards that impose implicit constraints on the acceptable error probabilities within a research community. They are solutions to a social‐epistemic problem of coordination among members of a community. They allow individual researchers in a community to develop a reliable sense of the dependability of certain kinds of scientific outcomes, such as the conclusions of epidemiological meta‐analyses, on the basis of their experience and their knowledge of the procedures, and without having personal knowledge of the person who has conducted the studies in question, her biases and motivations (Wilholt 2009,

98). That is to say, provided that a scientist has adequately followed the relevant conventions, other scientists may reliably estimate the reliability of her reported outcomes without knowing her personally.

On this view, in what sense are these values ‘arbitrary’? According to Wilholt,

‘the standards adopted are arbitrary in the sense that there could have been a different solution to the same coordination problem, but once a specific solution is socially adopted, it is in a certain sense binding’ (2009, 98). It is important to stress, though, that these values are arbitrary only to an extent and within a permissible range. The conventional critical p value could have arguably been six or four point six percent. Such values would have served as reasonable solutions to the community’s coordination problem. A critical p value of forty five percent, however, would not have worked, as it would have meant that the community would accept as statistically significant results that are just slightly higher than chance.

If we understand threshold values as social epistemic conventional solutions to a community’s coordination problem, how do social values fit into the picture? We can

‐ 193 ‐

identify two levels of influence, the level of the individual and the level of the community. As the level of the individual scientist, values may bias a scientist’s judgment in a way that will cause her to infringe an explicit or implicit conventional standard in a way that will increase the likelihood of arriving at her preferred result

(Wilholt 2009, 99). Put differently, values, such as personal or ideological investments in a certain outcome, may cause an individual scientist to lower or raise the evidence threshold in particular cases in a way that violates that community’s shared understanding of these thresholds. For example, a scientist who studies the efficacy of a certain drug and is funded by the pharmaceutical company that manufactures that drug may be motivated to reach an outcome that shows that the drug is efficacious. He may therefore lower the statistical threshold values required for acceptance in a way that violates the community’s conventions.

The phenomenon of motivated reasoning explains how this is psychologically possible. As the examples I will review in section 6 demonstrate, such biased judgment may occur at all kinds of stages in scientific work, such as experimental design, evidence characterization and interpretation of the results.

At the level of the community, values may influence the conventional threshold values themselves. For example, in the method of significance testing, there is an inherent mathematical trade‐off between minimizing false positives and minimizing false negatives.99 One cannot reduce the likelihood of the former without increasing the likelihood of the latter, and vice versa (Douglas 2008, 122). Values have influenced the balance that has been struck in practice between false positives and false negatives. The

99 False positives are the errors of accepting an experimental hypothesis as true when it is not. False negatives are the errors of rejecting an experimental hypothesis as false when it is not.

‐ 194 ‐

existing standards, which are manifested inter alia in the widespread choice of the five percent significance level, are clearly conservative in the sense that they regard false positives as more serious errors than false negatives (Wilholt 2009, 99). As Douglas argues, different social contexts may legitimately call for different balances between these two types of errors, which depend on the kinds of risks in the real world that may follow from wrongly rejecting a true hypothesis or wrongly rejecting a false hypothesis in a given context (Duoglas 2007, 123).

People’s different answers to the question of whether it is more important to prevent false positives or false negatives are not the only social values that may influence a community’s conventional choice of evidential threshold standards. Other social factors, such as rivalry with other fields, competition for funds, and the like may translate to values that may influence the evidential thresholds used in an epistemic community (cf. Bloor 1984). For example, in sociology, it is much harder to reach the levels of statistical significance that are reached in epidemiology. If sociologists used the same standards of statistical significance that are used in epidemiology, they would find it much harder to publish. Their desire to publish and the pressure they feel to do so may inter alia be responsible for their lower standards of statistical significance.

In some social contexts, scientists may adopt lower evidential thresholds while in other social settings they may adopt higher thresholds. For example, in a society where smoking is common, customary, and enjoyed by many, scientists may set higher thresholds of evidence about the dangers of smoking than in a society in which smoking is relatively rare and socially disapproved of. All other things being equal, the same evidence for the dangers of smoking may be considered insufficient in a smoking‐

‐ 195 ‐

friendly society and sufficient in a non‐smoking‐friendly society. And this may be so without a need arising in the smoking‐friendly society to come up with rival theoretical hypotheses that will explain away the evidence, as the underdetermination model requires.

Moreover, when social values and norms change, evidence thresholds may change accordingly. For example, if it becomes socially less acceptable to smoke for whatever reasons, the evidential thresholds may drop accordingly, again without ever a need arising for any theoretical justification for this drop. Therefore, when we explain changes of prevailing theories and beliefs over time, we must look at the chaining social values and their influence on the evidential thresholds required for belief in any given period.

5. Values Affect the Relative Weighing of Discordant Evidence

So far I have argued that values affect the trust we extend to the testimony of others and the thresholds we require our evidence to meet for us to justifiably believe or accept a theory. I have also argued that the role values play in these processes is not an aspect of filling the underdetermination gap, because when we decide that a speaker should not be trusted or that the evidence in support of a certain theory falls short of our evidential standards, we often do not, explicitly or implicitly, consider a rival theory that explains the evidence. Rather, we simply dismiss it.

In a similar way, values also affect our assessment of mixed evidence. Very often in science, we have multimodal evidence ‐ evidence from multiple techniques for the same hypothesis (Stegenga 2009, 651). Such evidence is often discordant. There are two types of discordance. The first is inconsistency – an apparent contradiction between the

‐ 196 ‐

hypotheses that evidence from different types supports. The second is incongruity – different results that have been produced under different background assumptions and are reported in ‘different languages’, namely using different and often incommensurable units (Stegenga 2009, 654). In the face of discordant evidence, one is naturally faced with the following questions: Which evidence is more relevant to the given case? What relative weight should be given to each type of evidence? Which evidence, among the discordant multimodal evidence, should be most trusted?

In science, there is no algorithmic and universally agreed upon method to combine multimodal evidence. A holistic appraisal of a body of evidence involves subjective judgment. In the same way values may lower or raise, to some extent, the thresholds we require our evidence to meet, they can also decrease or increase, to some extent, the relative weight an individual or a community assigns to each piece of evidence in a body of multimodal discordant evidence. Given the same body of evidence, different persons or different communities that adhere to different values which cause them to prefer one theory over another, may assign different relative weights to difference pieces of evidence.

The Bendectin trials, which I discussed in Chapter 4, provide an example of such influence. Recall that the question in dispute in these trials was whether the drug

Bendectin could cause birth defects in human embryos. Among the discordant evidence were structural activity studies, in vitro studies, in vivo studies, and epidemiological studies in humans, each with its own merits and drawbacks (see Table 3 on page 155).

While humans studies did not show correlation between exposure to Bendectin and birth defects, a minority of in vitro studies (dose‐response animal studies) did show

‐ 197 ‐

such a correlation, and the other types of evidence supported this possibility. The final outcome of the Bendectin litigation was the ruling that human epidemiological studies are required for establishing causation in humans, namely that human epidemiological studies carry the bulk of the evidential weight in settling the issue.

Also recall that Edmond and Mercer (2003) argue that it was not a priori determined that human studies carry most of the evidential weight in deciding this issue. Rather, this outcome was socially negotiated in the process of the litigation. How could social values influence this decision? The U.S. legal system is a jury system, where the jury – twelve citizens with no legal background – assume the role of the trial fact finder. The outcome of the Bendectin litigation was established in a period of growing concerns, especially by conservatives, about jurors’ abilities to carry out their role as fact finders in trials involving complex scientific evidence. These worries were raised in popular and academic monographs that received wide attention (e.g. Foster et al. 1993;

Huber 1993). It was argued that jurors tend to sympathize with the plaintiffs, who are typically ordinary working‐class citizens, and tend to be hostile to the defendant, typically a big corporation. Thus, they tend to decide in the plaintiffs’ favour based on

‘junk science’ or irrelevant evidence presented by the plaintiffs’ experts and purporting to show a causal link between the suspected substance and the plaintiffs’ illness. Such practice by the jury, it was argued, not only averts justice, but also inhibits economic development, as it compels big corporations to spend major funds in defending themselves in legal trials and paying compensation (Sanders 1998, Ch. 1 & 2) .

In light of these worries, it is plausible that the influence of social values, namely the expressed disapproval of the use of the legal system to extort big corporations and

‐ 198 ‐

to inhibit U.S. economic growth, would have been to raise the evidential standards and assign more relative weight to human epidemiological studies. By contrast, in the regulatory context, where the prevailing social values are preventing public health hazards, human studies are not required to establish causation. The International

Agency for Research on Cancer (IARC) and the U.S. National Toxicology Program, for instance, have identified at least five human carcinogens and about seventy‐five percent of probable human carcinogens on the basis of animal studies alone (Cranor 2005, 189).

The fact that different social values prevail in the legal and regulative contexts may explain the different weighing of discordant evidence in similar cases.

It is important to qualify the claim I have just made. In just the same way that social values may influence the raising and lowering of evidential thresholds only to an extent, so is their influence on the relative weighing of discordant evidence limited. The same psychological constraint that is associated with the phenomenon of motivated reasoning and requires people to be able to rationalize their beliefs operates here as well. In addition the practical constraints that are posed by the collective need of the standards to be effectively used as solutions to problems of community coordination apply here as well.

This means that while there is more than one seemingly rational way to weigh the same body of multimodal discordant evidence, not any weighing will do. For example, in the Bendectin controversy, while human studies did not show a correlation between exposure to Bendectin and birth defects, the minority of animal studies showed only a weak correlation and the rest did not show any correlation. It is plausible that if most of the animal studies had shown a strong correlation, for example, courts

‐ 199 ‐

would have faced more difficulties in dismissing them and assigning the bulk of the evidential weight to human studies. Nevertheless, as the comparison between the legal context and the regulatory context shows, the impact of social values is not necessarily inconsequential.

Against the picture I have just presented, one may argue that science can provide us with systematic and value‐free ways to combine and assess discordant multimodal evidence. Thus, the influence of social values on weighing multimodal discordant evidence may be eliminated. The prime example to support such a claim would be the hierarchy of evidence associated with the Evidence‐Based Medicine (EBM) movement, which is widely used in medicine. In a nutshell, according to the EBM evidence hierarchy, a randomized controlled trial (RCT) is better evidence than an observational study, and a meta‐analysis of RCTs is better evidence than a single RCT.

In order for this objection to work, however, it must be shown that the formulation of the EBM evidence hierarchy was not itself subject to the influence of social values, and that it rests on a sound independent epistemic rationale. This is not the case. As Mercer (2008) argues, the EBM movement, which gained momentum in the

1990s, was motivated by growing concerns about physicians’ over‐reliance on clinical experience and authoritative senior expert advice, rather than scientific clinical trials.

These concerns are parallel in important respects to the concerns that arose during the

Bendectin trials about the authority of experts and the so‐called unscientific nature of some of the evidence that was presented to courts.

More importantly, the EBM evidence hierarchy rests on a shaky epistemic basis.

While EBM may have all sorts of merits, such as improving critical thinking,

‐ 200 ‐

empowering young clinicians to challenge conventional wisdom, and generalizing research findings to large populations (Wilson 2010, 398‐9), these are social rather than epistemic merits. As philosophers have argued, arguments to epistemically justify the rigid EBM hierarchy of evidence do not hold up to philosophical scrutiny (Worrall

2002; 2007a; 2007b; Borgerson 2009; La Caze 2009, 511‐18).100 The only argument that stands opposed to such criticism is that randomization and blinding are an effective, though not the only possible method to reduce selection bias (Worrall 2007b,

1009; Borgerson 2009, 224). Hence, all other things being equal – which is rarely even approximately the case in practice – an RCT has better internal validity than an observational study (La Caze 2009, 516‐523). But the question of what properties of the evidence – internal validity, external validity or other properties – are more important in a given case is a value judgment that obviously depends on the social context.

Moreover, in practice, there is no one universally agreed upon EBM evidence hierarchy. Rather, there are several evidence hierarchies, which were produced by different official bodies. They all purport to implement the general principles of EBM.

These hierarchies weigh and rank different types of evidence differently, and may each lead to a different conclusion from the same body of evidence (Upshur 2003). Leaving aside the philosophical arguments against the epistemic rationales underpinning EBM, the de facto variety of EBM evidence hierarchies mitigates against the claim that medicine and science can provide us with algorithmic value‐free methods to combine and assess discordant evidence. This is because de facto, different researchers can legitimately use different evidence hierarchies, and thus their values may enter the

100 I could not find any sources that directly try to counter the epistemic arguments against the EBM hierarchy.

‐ 201 ‐

picture and influence the particular hierarchy they end up choosing. In particular, their directional goals and motivations to reach a certain outcome rather than another may cause them to choose the evidence hierarchy out of the available hierarchies that best favours the outcome for which they strive.

6. Applying the Framework to Existing Case Studies

6.1. Values and Evidence in Dioxin Cancer Research

My analysis sheds new light on some recently discussed examples in philosophy and sociology of science. Heather Douglas discusses a series of studies in which rats were exposed to dioxin and slides with their liver tissues were taken to determine if they had developed cancer. As part of the study, researchers needed to characterize the tissue samples. As it turns out, characterizing some slides, which is done by identifying certain visual patterns in them, is a subtle matter. Three different studies that used the same slides as data characterized some of them differently (Douglas 2000, 569‐72).

Douglas argues that values should not influence the characterization of clear evidence, but may and ought to influence the characterization of borderline evidence. In the example of the dioxin studies, clear cases of diseased tissues should be characterized as such and clear cases of healthy tissue should be characterized as such.

By contrast, so Douglas argues, all other things being equal, in a society mostly concerned with the dangers of cancer, borderline slide cases should be characterized as diseased, and in a society mostly concerned with the economic burden of overregulation, they should be characterized as healthy. Such a practice reflects the types and levels of inductive risk, i.e. the risk associated with accepting erroneous knowledge claims, that society is wiling to take (Douglas 2000, 569‐72; 2008a, 124).

‐ 202 ‐

This claim fits with a broader notion of knowledge as a concept that necessarily involves a pragmatic dimension.101

On Douglas’ account, however, it is not clear why the influence of social values is impermissible in one type of case but not in another. In cases such as the clear slides,

Douglas argues that values have no legitimate role to play in evidential reasoning because ‘our preferences for the world have no direct bearing on the way it actually is

[…] If our empirical reasoning were guided by such wishful thinking, we would have little chance of allowing the world to surprise us’ (2008b, 9‐10). It is not clear why the same argument does not apply to borderline evidence as well. By the same token, in the borderline slide case, just as in the clear slide case, a slide is either of a diseased tissue or a healthy tissue. Our characterization of it may get it right or may get it wrong, but it will not change the way the world actually is. Why, then, according to Douglas, is the influence of values epistemically legitimate in the borderline case?

My account may explain the rationale underpinning Douglas’ distinction. It is in the borderline cases that researchers are most prone to be engaged in motivated reasoning. In the borderline cases, researchers may be at much more liberty than in the clear cases to construct seemingly rational justifications for every characterization of the evidence they chose to furnish. All other things being equal, the less significant the visual pattern in the slide is, the more a researcher is prone to the influence of social values, and the less likely she is to be in an objective epistemic position to evaluate the evidence. If an implicit influence of a researcher’s idiosyncratic values may already be present and seems hard to avoid, we might as well make the influence of values on

101 Douglas (2000; 2008a) supports this view with considerations from scientific practice. For additional arguments to the same effect, see Stanley (2005) and Fantl & McGrath (2009).

‐ 203 ‐

evidence characterization explicit, recognize the values society deems important, and take them consciously into consideration when we evaluate the evidence. Or so an argument to justify Douglas’ normative distinction between permissible and impermissible cases of the influence of social values that makes use of my theoretical framework would run.

6.2. Values and Evidence in the Gravity Waves Controversy

The gravity waves controversy is another recently discussed case from physics on which my framework may shed new light. General Relativity predicts that moving massive bodies produce weak radiation‐like phenomenon known as ‘gravity waves’.

Collins (1981; 1985, 79‐112) followed a controversy in the community of physicists in the 1970s over the detection of gravity waves. Gravity waves are hard to detect and it was considered especially difficult to isolate their effects from that of other forms of radiation. The controversy concerned to a large extent the noise‐signal distinction. It centred on the question of whether physicist Joseph Weber had managed to build a working gravity‐wave detector, or whether his detector was producing ‘pure noise’.

While in the early 1970s Weber’s claim to have successfully detected gravity waves was viewed as having some credibility, toward the end of the decade a consensus emerged that his claim was incorrect. Among the claims Weber’s critics made were that his detected values were significantly higher than the theory predicted, that there were difficulties in replication, that his computer analysis of the data was flawed and that he had found a correlation between two isolated detectors which were later discovered to be asynchronous and hence could not possibly detect the same events (Collins 1981, 38‐

44).

‐ 204 ‐

Despite that, Collins argues that evidence alone could not settle the controversy.

Social values were required as well. Collins argues that Weber had seemingly rational replies to all the criticism and that none of the physicists Collins interviewed found all of the criticism persuasive (1981, 49‐54). In response, Allan Franklin (1994) argues that scientists’ rejection of Weber’s claims was reasoned and rational and hence no social values were needed to resolve the controversy. While Franklin may be right to argue that the final outcome of the controversy was rational,102 he does not manage to show that it was not contingent on its social settings. Franklin’s account does not disprove

Collin’s claim that Weber had seemingly rational replies to his critics, and that for all of his claims, some of the critics found some of his replies reasonable. Had more members of the scientific community been as motivated as Weber, for whatever reason, to succeed in detecting gravity waves using his detector, a different consensus might have emerged.103 Because the controversy centred on the adequate evidential threshold required to justifiably make a detection claim, social values inevitably played a role in bringing the controversy to closure, irrespective of whether the final outcome was epistemically justified or not.

Almassi argues that both Weber and his critics could be simultaneously rationally justified in their respective contrary beliefs. The general lesson he draws from this example is that rational disagreement between epistemic peers is possible

(Almassi 2009a, 581).104 If this were a reasonable disagreement where both sides were

102 In fact, the consensus that emerged seems epistemically justified, as it prima facie meets the three conditions I have listed in Chapter 3 of a knowledge‐based consensus. 103 This, of course, does mean that it would have necessarily lasted, as physicists might have reasonably expected detectors to improve over time. 104 See Christensen (2009) for an overview of the different positions and arguments in the debate about the possibility of rational peer disagreement.

‐ 205 ‐

rational, one might argue that any outcome of the dispute might have been rational. This seems to be Collins’ own position. Otherwise, it would be hard to explain how, after having argued that the consensual outcome of the controversy in the gravity waves case was contingent on social factors, Collins can argue from a normative perspective that in esoteric fields such as gravity wave research, the relevant community of researchers has the epistemic right to settle disputes by forming a consensus (Collins & Evans 2002,

242‐3).

My analysis militates against regarding this as a case of reasonable disagreement, and hence against Collin’s view that any resolution of controversy in an esoteric community of researchers is epistemically justified. With respect to the critics, as I argued in Chapter 1, social evidence about one’s credibility is typically insufficient for epistemic justification. With respect to Weber, the phenomenon of motivated reasoning casts doubt on Weber’s ability to objectively and rationally reason about his own data, especially when they were borderline between signal and noise. As I have repeatedly argued throughout this thesis, a proper epistemic evaluation of the controversy needs to be done at the level of the community rather than the individuals of whom it consists.

What these case studies show is that any depiction of scientists either as strictly following the evidence and reaching warranted conclusions based on it, or as socially negotiating and constructing facts based solely on their interests, values and goals, irrespective of the evidence they have is probably mistaken. But my point has not been to argue for a middle‐ground position between social constructivist and a rationalist account of science. Rather, it has been to point out the need to pay close attention to the

‐ 206 ‐

different values that operate and their influence of the threshold standards and the relative weighing of discordant evidence, to see how they change over time and how their effects differ in different social contexts in which similar issues are considered. As

I will argue in the next section, my framework has implications not only to the analysis of particular case studies, but also for theories of knowledge, truth, and justification in epistemology and philosophy of science in general.

7. Broader Implications for Epistemology and Philosophy of Science

I would like to briefly address the significance of my epistemic framework in a broader epistemic context. Evidence plays an important part in epistemology. For example,

Evidentialism, a prominent normative theory of epistemic justification, states that a person ought always to base her beliefs solely on her evidence (Conee & Feldman

2004). The worry to which such studies give rise is that if people do not and perhaps cannot reason rationally, then evidentialism imposes on people an epistemic duty they cannot perform, thus violating the principle that ‘ought’ implies ‘can’. Conee and

Feldman reply to this worry inter alia by trying to cushion the blow of such studies, and by interpreting their results in ways that do not necessarily entail that people reason irrationally (Feldman 2003, 162‐5; Conee & Feldman 2004, 86‐7). The problem is that

Feldman and Conee only discuss ‘cold’ cognitive fallacies such as incorrect use of the rules of probability or generalizing from unrepresentative data. They do not address

‘hot’ cognitive biases, such as motivated reasoning, which problematize the status of evidentialism as an adequate normative theory of justification. It seems to follow from my argument that often individuals and communities are unable to follow the evidence

‐ 207 ‐

without being influenced by values, which poses a prima facie problem for evidentialism as a normative theory of justification.

Such studies pose a threat also to Goldman’s Veristic Social Epistemology, which takes the task of social epistemology to be identifying truth‐yielding social‐epistemic practices. One such practice, according to Goldman, is deference to science. Goldman interprets social constructivist sociological studies of science as arguing that science does not produce truths, but rather reflects the political interests of various actors.

Goldman admits that the studies that show the existence of motivational reasoning are a possible support to constructivists’ claims, as they may be used to show that scientists reason in ways that reflect their interests. Rather than seeing the glass half empty,

Goldman sees it half full. He sees it as a positive sign that people cannot just believe anything they want, but rather only what they can rationally justify to themselves. He also expresses optimism about science’s capacity for self‐correction. He points out female scientists’ success in correcting male gender bias in primatology as an example supporting his optimism (Goldman 1999, 234‐8).

There are reasons to doubt, however, whether Goldman’s optimism is justified.

First, recall that as Kunda stresses, although people are restricted in their motivated reasoning by their ability to construct a justification for their beliefs, ‘the objectivity of this justification construction is illusory because people do not realize that the process is biased by their goals’ (Kunda 1990, 483; emphasis added). Biases in science are a good reason for concern. Biases of which scientists are unaware are even a better reason.

Second, in cases such as the ones Goldman describes, the change in scientific views was triggered ‘from outside’ science. In the case of primate research, it was

‐ 208 ‐

triggered by feminist activism and the rise in women’s social status. Only when the social conditions were such that women entered the science of primatology, were the existing dogmas challenged. Other examples are available. Until recently, it has been very common in medicine to regard males as the ‘standard patients’, and to exclude women from clinical trials, on the pretence that their menstruation is a confounding factor that interferes with tests. This practice resulted in absent or inaccurate medical knowledge of the effect of drugs on women. Only in the last decade have things started to change, and women have started to be regularly included in clinical trials. In this case as well, the change was not initiated from within scientific circles (Arber & Thomas

2001). In AIDS research during the 1980s and 1990s, epistemically‐positive changes in protocols of trials testing the efficacy of drugs were initiated, despite initial objections by the medical establishment, following demands by AIDS patients activists (Epstein

1996). What is unique about this example is the activists’ social status – young educated white males who had enough power to bring about these changes. Other groups are typically much weaker. If society moves toward greater social justice at all, it does so very slowly, two steps forward, one step back. If the capacity of science for correcting its errors depends on outside social pressure rather than on science itself, there seem be more reasons for pessimism than optimism.

My aim is to argue neither for pessimism about scientific progress nor for the inadequacy of evidentialism as a normative theory of justification. Rather, my point is that Goldman, Conee and Feldman’s responses to such worries are unsatisfactory. In cases in which the evidence is clear‐cut, the recommendation of evidentialism to form beliefs based on the evidence and the recommendation of veristic social epistemology

‐ 209 ‐

to trust science as the most reliable truth‐yielding social institution are trivial. These recommendations become interesting and disputable only when the evidence is less than perfect and belief is potentially value‐laden. If we are to examine the prospects of evidentialism as a normative theory of justification, or the prospects of veristic social epistemology as a realistic enterprise, my framework must be taken into consideration.

Conclusion

In this chapter, I have identified three roles social values play in evidential justification in addition to their familiar role as filling the logical gap of underdetermination between theory and evidence. First, values affect trust in testimony. Second, values lower and raise evidential thresholds. Third, values affect the relative weighing of discordant evidence.

Drawing on research in experimental psychology, I have argued that while the effect of values on evidential reasoning in the ways I have identified is constrained by agents’ abilities to rationalize their beliefs, it is nevertheless not trivial or insignificant.

The social context and the values that prevail in it do make a difference in theory acceptance and belief formation.

Consequently, I have argued that any attempt by philosophers and historians of science and STS scholars to account for differences in beliefs or theories in different periods or different social contexts must take into account the relevant evidence, social values and the ways they can interact according to my model. Similarly, any normative epistemic theory in philosophy of science or epistemology about the role of values in epistemic justification must take my descriptive model into account as well.

‐ 210 ‐

Conclusion

One of our major aims in life is having knowledge, either for its own sake, or in order to take informed actions on issues varying from the mundane to the catastrophic. One of the best, common, and in many cases only ways to have knowledge is through social interaction. In this thesis, I developed a social theoretical framework for dealing with knowledge that can help us tell when this aim is satisfied – when we have managed to acquire knowledge, as opposed to mere belief or opinion.

The starting point of my thesis was a study of the aspects of the concept of knowledge that are relevant to my project. Analytic epistemology offers many competing conceptions of knowledge. Their common denominator is that they are nearly all individualistic. They take knowledge to be the property of individual agents.

They seem to assume the picture of the isolated autonomous epistemic agent that was conceived in the Enlightenment period. The rationalist tradition would imagine such an agent sittings alone in his armchair and acquiring knowledge by reasoning from first principles. The empiricist tradition would imagine him being born into the world as a tabula rasa acquiring knowledge from experience and perception. On both these images, however, he single‐handedly acquires justification for the beliefs he forms.

Actual knowers, especially in our day and age, are different from this picture.

They acquire most of their beliefs from the testimony of others, including experts, and from social institutions, such as science, that are in charge of the generation of knowledge. While social epistemology has devoted much attention to the study of testimony as a source of knowledge, it still construes knowledge in individualistic terms. I have argued that if our conceptions of knowledge are to acknowledge the depth

‐ 211 ‐

of our epistemic dependency and reliance on others, we must abandon epistemic individualism and adopt epistemic communalism instead.

On the picture I have proposed, a normal adult individual does not possess justification for most of her beliefs that are ordinarily called ‘knowledge’. Rather, other members of her community, with whom she forms relations of trust, possess the justification for most of her beliefs. If we think of justification in terms of having evidence, then most of a person’s beliefs are justified by evidence other people in her community have. If we think of justification in terms of a reliable belief forming process, then the process extends far beyond a person’s cognition into the community to which she belongs, where the salient justificatory work is typically done. I have argued that attempts by social epistemologists to resist this picture ultimately fail.

If analytic epistemology is too individualistic and does not take seriously enough the social dimensions of knowledge, contemporary SSK is at another extreme. It wishes to subsume the notion of knowledge under the social, and do away with epistemic notions such as justification and rational belief. Under the SSK conception of knowledge, knowledge is merely or mainly social agreement between persons that aims at promoting their individual or collective interests. I have argued that such a conception of knowledge is inadequate. I have presented the PRIMES affair, in which scientists infringed the norms of peer‐reviewed scholarly publication, and communicated a distorted account of their discovery to the general press in order to gain public visibility and win a possible priority race. Despite that, they managed to get recognition for their discovery. They were not penalized by their peers for infringing the community norms, and their peers could grasp the meaning of their discovery through press‐distorted

‐ 212 ‐

accounts of it. I have argued that if we regard knowledge as mere social agreement, this outcome is surprising and hard to explain. By contrast, if we assume that there are epistemic standards that are independent of the social context, and that scientists could appeal to them in this case, the outcome of PRIMES becomes easily explainable. As I have argued, there is nothing unique about PRIMES with respect to other case studies in

SSK that should make its outcome difficult for SSK theory to explain. Hence, the fact that

SSK theory faces difficulties with explaining the outcome of PRIMES exposes a general difficulty with the SSK conception of knowledge as mere social agreement.

If knowledge is a communal good, but is not mere social agreement, then what distinguishes between them? I have addressed this question at two levels. At the conceptual level, I have identified four types of social agreement or consensus, three of which are not knowledge based and one is. The first type of consensus is a non‐ epistemic consensus. This is a consensus that aims at promoting non‐epistemic aims.

For example, experts may mask internal disagreements among them and present a unified front in order to promote a non‐epistemic aim they share, such as securing research budgets. The second type of consensus that is not knowledge based is a vertically‐lucky consensus. This is an agreement that happens to be correct, but could have easily been wrong. It is just a matter of luck – a rare arrangement of contingent circumstances – that the parties to the agreement reached a correct view. The third type of a non‐knowledge‐based consensus is an epistemically unfortunate consensus, in which the parties to the consensus have the bad luck of being systematically or deliberately mislead. For example, they may all share a bias, which they find hard to

‐ 213 ‐

transcend, that causes them to reason incorrectly based on the evidence. When a consensus belongs to none of these three types, it is likely to be knowledge based.

In accordance with the aim of this thesis, which is assisting us in telling when we have successfully acquired knowledge, I have translated this conceptual analysis into evaluative criteria that can tell us, in all likelihood, when a particular consensus is knowledge based. The connecting link between the conceptual analysis and my suggested evaluative criteria is inference to the best explanation. I have argued that when we can eliminate non epistemic factors, veritic epistemic luck, and epistemic misfortune as good explanations of a consensus, knowledge remains its best explanation, and hence we can legitimately infer that the consensus is knowledge based.

I have identified three conditions which are jointly sufficient for knowledge being the best explanation of a consensus. First, when a consensus exhibits an apparent consilience of different lines of evidence, namely when it seems to be built on an array of evidence that is drawn from a variety of techniques and methods, it is less likely to be an accidental by‐product of one technique – and all the more likely to be knowledge based. Second, when a consensus is socially diverse, i.e. shared by men and women, researchers from the private and public sectors, liberals and conservatives, it is more likely to be knowledge‐based. Third, there must be meta‐agreement, namely the agreement must be genuine. Scientists must give the same meaning to the same terms and share the same fundamental background assumptions.

I have demonstrated that when applied to a concrete case study, my theory of knowledge‐based consensus leads to non‐trivial conclusions. In the controversy in U.S. courts about whether the drug Bendectin could cause birth defects, courts justified their

‐ 214 ‐

decision that it could not inter alia by deferring to scientific consensus. An application of my theory to this consensus, however, shows that it was not knowledge based. I have therefore argued against the common practice in policy and law of naive deference to consensus in order to resolve disputes. In order to be justified, any deference to consensus must evaluate the consensus by the criteria my theory specifies.

Last, I have proposed a new theory of the logical relations between social values and evidential justification, which, as I have argued, are not exhausted by the existing model of values as filling the logical gap of underdetermination of theory by evidence. I have identified three ways in which values and evidence are interrelated. First, values affect the trust we extend to the testimony of others. In light of my analysis of knowledge as a communal good that depends on irreducible trust between members of an epistemic community, this should not be surprising. Second, values influence the threshold values that determine when a theory is accepted and rejected. Such influence occurs both at the level of the epistemic community and at the level of the individual scientist. Third, values influence the process of combining different lines of evidence. As there is no single algorithmic way to do that, different people can give different weights to different evidence based on the values they cherish. I have argued that any evaluation of our theories and beliefs and the extent to which they are supported by evidence must closely examine the ways in which social values have affected our evidential reasoning.

My account of the relations between values and evidence together with my theory of knowledge‐based consensus constitute a potent tool that can help us evaluate many of our theories and beliefs and tell whether and when they amount to knowledge.

I believe that further work that is based on my theory has the potential to make

‐ 215 ‐

significant contributions to the relevant scholarly discourse. First, my social theory of knowledge can be applied to a variety of case studies in order to determine whether the theories and beliefs in them amount to knowledge. In turn, since every case study has its own complexities, such case studies may be used to further refine and nuance my theory. Second, my approach may be expanded to deal with cases of partial consensus, dissent, and disagreement as well as consensus. In the same manner that we can find indicators that tell us when knowledge is the best explanation of a consensus, we can find general indicators for knowledge being the best explanation of dissent as well. That is, we can ask under what conditions knowledge is the best explanation of dissent, or particular types of dissent. Third, my theory may have significant implications for existing debates in epistemology about the nature of knowledge and justification. For example, according to evidentialism, a person ought to form beliefs based only on the evidence she has. However, if we accept the principle that ‘ought’ implies ‘can’, and my claim that in some circumstance people cannot avoid being influenced by values when they engage in evidential reasoning, it might follow that evidentialism is incorrect, and an argument might be made in support views that connect knowledge and practical interests. I hope that this is just the tip of the iceberg, and that my theory will prove itself useful in the fields of social epistemology, philosophy of science, and STS.

‐ 216 ‐

Bibliography

Adler, Emanuel. 2005. Communitarian International Relations: The Epistemic Foundations of International Relations. London: Routledge.

Adler, Jonathan E. 1994. Testimony, Trust, Knowing. The Journal of Philosophy 91(5): 264‐275.

Adler, Jonathan E. 2007. Argumentation and Distortion. Episteme: A Journal of Social Epistemology 4(3): 382‐401.

Agrawal, Manindra, Neeraj Kayal & Nitin Saxena. 2002. PRIMES is in P. .

Agrawal, Manindra, Neeraj Kayal & Nitin Saxena. 2004. PRIMES is in P. Annals of Mathematics 160(2): 781‐93.

Alatout, Samer. 2009. Bringing Abundance into Environmental Politics: Constructing a Zionist Network of Water Abundance, Immigration, and Colonization. Social Studies of Science 39(3): 363‐394.

Almassi, Ben. 2007. Experts, Evidence, and Epistemic Independence. Spontaneous Generations: A Journal for the History and Philosophy of Science 1(1): 58‐66.

Almassi, Ben. 2009a. Conflicting Expert Testimony and the Search for Gravitational Waves. Philosophy of Science 76(5): 570‐584.

Almassi, Ben. 2009b. Trust in Expert Testimony: Eddington’s 1919 Eclipse Expedition and the British Response to General Relativity. Studies in History and Philosophy of Modern Physics 40(1):57–67.

Angell, Marcia. 2009. Drug Companies and Doctors: A Story of Corruption. The New York Review of Books 56(1): 8‐13.

Anonymous. 2002. New Algorithm by Three Indians. The Hindu (9 August) .

Anonymous. 2003. Mathematicians are Discussing Ways to Make Code‐Breaking Easier. The Economist 366(8317) (29 March): 89.

Arber, Sara & Hilary Thomas. 2001. From Women’s Health to a Gender Analysis of Health. In The Blackwell Companion to Medical Sociology, ed. William C. Cockerham, 94‐113. Malden, MA: Blackwell.

‐ 217 ‐

Audi, Robert. 1997. The Place of Testimony in the Fabric of Knowledge and Justification. American Philosophical Quarterly 34(4): 405‐422.

Baier, Annette. 1986. Trust and Antitrust. Ethics 96(2): 231‐260.

Baigrie, Brian S. & Jagdish. N. Hattiangadi. 1992. On Consensus and Stability in Science. The British Journal for the Philosophy of Science 43(4): 435‐458.

Barnes, Barry & David Bloor. 1982. Relativism, Rationalism and the Sociology of Knowledge. In Rationality and Relativism, ed. Martin Hollis & Steve Lukes, 21‐ 47. Oxford: Blackwell.

Beatty, John. 2006. Masking Disagreement among Experts. Episteme: A Journal of Social Epistemology 3(1): 52‐67

Beecher‐Monas, Erica. 2002. Evaluating Scientific Evidence: An Interdisciplinary Framework for Intellectual due Process. Cambridge: Cambridge University Press.

Ben Menahem, Yemima. 1990. Inference to the Best Explanation. Erkenntnis 33: 319‐ 334.

Blaauw, Martijn and Duncan Pritchard. 2005. Epistemology A­Z. New York: Palgrave Macmillan.

Bloor, David. 1984. A Sociological Theory of Objectivity. In Objectivity and Cultural Divergence, ed. S. C. Brown, 229‐45. Cambridge: Cambridge University Press.

Bloor, David. 1991. Knowledge and Social Imagery, 2nd ed. Chicago: University of Chicago Press.

Bloor, David. 1999. Anti‐Latour. Studies in History and Philosophy of Science 30(1): 81‐ 112.

Borgerson, Kirstin. 2009. Valuing Evidence Bias and the Evidence Hierarchy of Evidence‐Based Medicine. Perspectives in Biology and Medicine 52(2): 218‐233.

Bornemann, Folkmar. 2003. PRIMES is in P: A Breakthrough for “Everyman”. Notices of the American Mathematical Society, 50(5): 545‐52.

Bowler, Peter J. & Iwan R. Morus. 2005. Making Modern Science: A Historical Survey. Chicago: University of Chicago Press.

Brent, Robert. 1985. Editorial Comment on Comments on “Teratogen Update: Bendectin”. Teratology 31(3): 429‐430.

Brizon, Uriel (2002) ‘The Prime Numbers will be Identified, the Code Will Be Broken’, Haaretz (19 August): 6. [Hebrew].

‐ 218 ‐

Broks, Peter. 2006. Understanding Popular Science. Maidenhead: Open University Press.

Brown, James R. 1989. The Rational and the Social. London: Routledge.

Brown, James R. 1994. Smoke and Mirrors: How Science Reflects Reality. London: Routledge.

Brown, James R. 2001. Who Rules in Science? An Opinionated Guide to the Wars. Cambridge, MA: Harvard University Press.

Brown, James R. 2008. The Community of Science®. In The Challenge of the Social and the Pressure of Practice: Science and Values Revisited, eds. Martin Carrier, Don Howard & Janet Kourany, 189‐216. Pittsburgh, PA: University of Pittsburgh Press.

Brown, Kenneth S. et al. 1985. Comments on “Teratogen Update: Bendectin”. Teratology 31(3): 431‐431.

Bryant, Arthur H. & Alexander Reinert. 2001. Epidemiology in the Legal Arena and the Search for Truth. American Journal of Epidemiology 154(12) Supp.: S27‐S35.

Bucchi, Massimiano. 1996. When Scientists Turn to the Public: Alternative Routes in Scientific Communication. Public Understanding of Science 5: 375‐94.

Callon, Michel & Bruno Latour. 1992. Don’t Throw the Baby Out with the Bath School! A Reply to Collins and Yearley. In Science as Practice and Culture, ed. Andrew Pickering, 327‐342. Chicago: The University of Chicago Press.

Christensen, David. 2009. Disagreement as Evidence: The Epistemology of Controversy. Philosophy Compass 4(5): 756–767.

Coady, C. A. J. 1994. Testimony, Observation and ‘Autonomous Knowledge’. In Knowing From Words: Western and Indian Philosophical Analysis of Understanding and Testimony, eds. Bimal Krishna Matilal & Arindam Chakrabarti, 225‐250. Dordrecht: Kluwer.

Cohen, Jonathan L. 1992. An Essay on Belief and Acceptance. Oxford: Clarendon Press.

Collins, Harry M. & Robert Evans. 2002. The Third Wave of Science Studies: Studies of Expertise and Experience. Social Studies of Science 32(2): 235‐96.

Collins, Harry M. & Trevor J. Pinch. 1993. The Golem: What Everyone Should Know about Science. Cambridge: Cambridge University Press.

Collins, Harry M. 1981. Son of Seven Sexes: The Social Destruction of a Physical Phenomenon. Social Studies of Science 11(1): 33‐62.

‐ 219 ‐

Collins, Harry M. 1982. The Replication of Experiments in Physics. In Science in Context: Readings in the Sociology of Science, ed. Barry Barnes and David Edge, 94‐117. Cambridge, MA: MIT Press.

Collins, Harry M. 1983. The Sociology of Scientific Knowledge: Studies of Contemporary Science. Annual Review of Sociology 9: 265‐285.

Collins, Harry M. 1985. Changing Order: Replication and Induction in Scientific Practice. Chicago: The University of Chicago Press.

Collins, Harry M. 2010. Humans not Instruments. Spontaneous Generations: A Journal for the History and Philosophy of Science 4(1): 138‐47.

Colyvan, Mark. 2008. Indispensability Arguments in the Philosophy of Mathematics. The Stanford Encyclopedia of Philosophy (Fall 2008 Edition), ed. Edward N. Zalta. http://plato.stanford.edu/archives/fall2008/entries/mathphil‐indis/.

Comesana, Juan. Forthcoming. Evidentialist Reliabilism. Noûs.

Conee, Earl & Richard Feldman. 2004. Evidentialism: Essays in Epistemology. New York: Oxford University Press.

Corlett, Angelo J. 1996. Analyzing Social Knowledge. Lanham, MD: Rowman & Littlefield.

Corlett, Angelo J. 2007. Analyzing Social Knowledge. Social Epistemology 21(3): 231‐ 247.

Cormen , Thomas et al. 2001. Introduction to Algorithms, 2nd ed. Cambridge, MA: MIT Press.

Cranor, Carl F. 2005. The Science Veil over Tort Law Policy: How Should Scientific Evidence be Utilized in Toxic Tort Law. Law and Philosophy 24(2): 139‐210.

Dancy, Jonathan. 1985. An Introduction to Contemporary Epistemology. Oxford: Blackwell.

Darwin, Charles. 1871. The Descent of Man, and Selection in Relation to Sex. London: John Murray.

Dascal, Marcelo. 1998. The Study of Controversies and the Theory and History of Science. Science in Context 11(2): 147‐154.

Daston, Lorraine. 2009. Science Studies and the History of Science. Critical Inquiry 35: 798‐815.

Daubert v. Merrell Dow Pharmaceuticals, Inc. 1993. 509 U.S. 579.

‐ 220 ‐

Day, Timothy & Harold Kincaid. 1994. Putting Inference to the Best Explanation in Its Place. Synthese 98(2): 271‐295.

DeRose, Keith. 1995. Solving the Skeptial Problem. Philosophical Review 111: 167‐203.

Douglas, Heather. 2000. Inductive Risk and Values in Science. Philosophy of Science 67(4): 559‐579.

Douglas, Heather. 2008a. Science, Policy, and the Value­Free Ideal. Pittsburgh, PA: University of Pittsburgh Press.

Douglas, Heather. 2008b. The Role of Values in Expert Reasoning. Public Affairs Quarterly 22(1): 1‐18.

Duesberg, Peter. 1996. Inventing the AIDS Virus. Washington, D.C.: Regnery Publishing.

Edmond, Gary & David Mercer. 2000. Litigation Life: Law‐Science Knowledge Construction in (Bendectin) Mass Toxic Tort Litigation. Social Studies of Science 30(2): 265‐316.

Enoch, David. 2009. Not Just a Truthometer: Taking Oneself Seriously (but not Too Seriously) in Cases of Peer Disagreement. http://law.huji.ac.il/upload/truthometer.pdf

Epstein, Steven. 1996. Impure Science: AIDS, Activism, and the Politics of Knowledge. Berkeley: University of California Press.

Fantl, Jeremy & Matthew McGrath. 2009. Knowledge in an Uncertain World. Oxford: Oxford University Press.

Feldman, Richard. 2003. Epistemology. Upper Saddle River, NJ: Prentice Hall.

Feldman, Richard. 2006. Epistemological Puzzles about Disagreement. In Epistemology Futures, ed. S. Hetherington, 216‐236. Ox ford: Oxford University Press.

Feyerabend, Paul. 1970. Consolations for the Specialist. In Criticism and the Growth of Knowledge, eds. Imre Lakatos and Alan Musgrave. Cambridge: Cambridge University Press.

Foster, Kenneth R., David E. Bernstein & Peter W. Huber. 1993. Phantom Risk: Scientific Inference and the Law. Cambridge, MA: MIT Press.

Fox, Robert. 1971. The Caloric Theory of Gases: From Lavoisier to Regnault. Oxford: Clarendon Press.

Frances, Bryan. 2005. Scepticism Comes Alive. New York: Oxford University Press.

‐ 221 ‐

Franklin, Allan. 1994. How to Avoid the Experimenters’ Regress. Studies in History and Philosophy of Science 25(1): 463‐491.

Fricker, Elizabeth. 1995. Telling and Trusting: Reductionism and Anti‐Reductionism in the Epistemology of Testimony. Mind 101(414): 393‐411.

Fricker, Elizabeth. 2002. Trusting Others in the Sciences: a priori or Empirical Warrant? Studies in History and Philosophy of Science 33(2): 373‐383.

Fricker, Miranda. 1998. Rational Authority and Social Power: Towards a Truly Social Epistemology. Proceedings of the Aristotelian Society 98: 159‐177.

Fricker, Miranda. 2007. Epistemic Injustice: Power and the Ethics of Knowing. Oxford: Oxford University Press.

Frye v. United States. 1923. 293 F.2d 1013 (D.C. Cir.)

Fuller, Steve. 1997. Science. Buckingham: Open University Press.

Fuller, Steve. 2002. Social Epistemology, 2nd ed. Bloomington, IN: Indiana University Press.

Fuller, Steve. 2007. The Knowledge Book: Key Concepts in Philosophy, Science and Culture. Montreal: McGill‐Queen's University Press.

Galen. 1963. On the Passions and Errors of the Soul, trans. Paul W. Harkins. Columbus, OH: Ohio State University Press.

Galison, Peter. 1997. Image and Logic: A Material Culure of Microphysics. Chicago: University of Chicago Press.

Gettier, Edmund. 1963. Is Justified True Belief Knowledge? Analysis 23: 121‐123.

Gibeault, Amanda. 2008. Knowledge of Witchcraft and Naturalized Epistemology. Talk given at New Directions in Epistemology: International Symposium, Canadian Society for Epistemology, Carleton University, November 21.

Giere, Ronald N. & Barton Moffatt. 2003. Distributed Cognition: Where the Cognitive and the Social Merge. Social Studies of Science 33(2): 301‐10

Giere, Ronald N. 1988. Explaining Science: A Cognitive Approach. Chicago: University of Chicago Press.

Giere, Ronald N. 2007. Distributed Cognition without Distributed Knowing. Social Epistemology 21(3): 313‐320.

Gilbert, Margaret. 1987. Modeling Collective Belief. Synthese 73(1): 185–204.

‐ 222 ‐

Gingras, Yves. 1995. Following Scientists through Society? Yes, but at Arm’s Length! In Scientific Practice: Theories and Stories of Doing Physics, ed. Jed Z. Buchwald, 123‐150. Chicago: University of Chicago Press.

Goldberg, Sanford C. 2007. Anti­Individualism: Mind and Language, Knowledge and Justification. Cambridge: Cambridge University Press.

Goldberg, Sanford C. 2010. Relying on Others: An Essay in Epistemology. New York: Oxford University Press.

Goldman, Alvin I. 1994. Psychological, Social, and Epistemic Factors in the Theory of Science. PSA 2: 277‐286.

Goldman, Alvin I. 1999. Knowledge in a Social World. New York: Oxford University Press.

Goldman, Alvin I. 2001. Experts: Which Ones Should You Trust? Philosophy and Phenomenological Research 63(1): 85‐111.

Goldman, Alvin I. 2002a. What Is Social Epistemology? A Smorgasbord of P rojects. In Pathways to Knowledge: Private and Public, 182‐204. New York: Oxford University Press.

Goldman, Alvin I. 2002b. Knowledge and Social Norms. Science 296 (June 21): 2148‐9.

Goldman, Alvin I. 2008. Immediate Justification and Process Reliabilism. In Epistemology – New Essays, ed. Quentin Smith, 63‐82. Oxford: Oxford University Press.

Goldman, Alvin I. 2009. Reliabilism. In The Stanford Encyclopedia of Philosophy, Fall 2009 Edition, ed. Edward N. Zalta. http://plato.stanford.edu/archives/fall2009/entries/ reliabilism/.

Goldman, Alvin I. Forthcoming. Toward a Synthesis of Reliabilism and Evidentialism? Or: Evidentialism’s Problems, Reliabilism’s Rescue Package. In Evidentialism and Its Discontents, ed. T. Dougherty. Oxford: Oxford University Press. http://alturl.com/m8j8

Gomes, Lee. 2002. A Beautiful Mind from India is Putting the Internet on Alert. The Wall Street Journal (4 November): B1.

Goodstein, Laurie & Greg Myre. 2005. Clerics Fighting a Gay Festival for Jerusalem. The New York Times (March 31). http://www.nytimes.com/2005/03/31/international/ worldspecial/31gay.html.

Gould, Stephen J. 1996. The Mismeasure of Man, rev ed. New York: Norton.

Grasswick, Heidi. 2008. Feminist Social Epistemology. In The Stanford Encyclopedia of Philosophy, Fall 2008 Edition, ed. Edward N. Zalta.

‐ 223 ‐

http://plato.stanford.edu/archives/fall2008/entries/feminist‐social‐ epistemology/.

Green, Leslie. 1991. Two Views of Collective Rights. The Canadian Journal of Law and Jurisprudence 4(2): 315–327.

Green, Michael D. 1996. Bendectin and Birth Defects: The Challenge of Mass Toxic Substances Litigation. Philadelphia: University of Pennsylvania Press.

Gregory, Jane & Steve Miller. 1998. Science in Public: Communication, Culture, and Credibility. New York: Plenum Press.

Grinnel, Frederick. 1999. Ambiguity, Trust, and the Responsible Conduct of Research. Science and Engineering Ethics 5(2): 205‐14.

Haack, Susan. 1998. Manifesto of a Passionate Moderate: Unfashionable Essays. Chicago: University of Chicago Press.

Haack, Susan. 2003. Entangled in the Bramble‐Bush: Science and the Law. In Defending Science – within Reason: Between Scientism and Cynicism, 233‐264. Amherst, NY: Prometheus Books.

Haack, Susan. 2004. Truth and Justice, Inquiry and Advocacy, Science and Law. Ratio Juris: An International Journal of Jurisprudence and Philosophy of Law 17(1):15‐ 26.

Haack, Susan. 2005. Trial and Error: The Supreme Court's Philosophy of Science. American Journal of Public Health 95(S1): S66‐S73.

Habermas, Jürgen. 1984. The Theory of Communicative Action, Vol. 1 & 2. Boston: Beacon Press.

Hackett, Edward J., Olga Amsterdamska, Michael Lynch, & Judy Wajcman, eds. 2007. The Handbook of Science and Technology Studies, 3rd ed. Cambridge, MA: MIT Press.

Hacking, Ian. 1983. Representing and Intervening: Introductory Topics in the Philosophy of Natural Science. Cambridge: Cambridge University Press.

Hacking, Ian. 1992. Statistical Language, Statistical Truth and Statistical Reason: The Self‐Authentication of a Style of Scientific Reasoning. In The Social Dimension of Science, ed. Ernan McMullin, 130‐157. Notre Dame, IN: University of Notre Dame Press.

Hacking, Ian. 1999. The Social Construction of What? Cambridge, MA: Harvard University Press.

Hacking, Ian. 2002. Historical Ontology. Cambridge, MA: Harvard University Press.

‐ 224 ‐

Hagstrom, Warren O. 1965. The Scientific Community. New York: Basic Books.

Halfon, Saul. 2006. The Disunity of Consensus: International Policy Coordination as Socio‐Technical Practice. Social Studies of Science 36(5): 783‐807.

Harding, Sandra. 2002. Must the Advance of Science Advance Global Inequality? The International Studies Review 4(2): 87‐105.

Hardwig, John. 1985. Epistemic Dependence. The Journal of Philosophy 82(7): 335‐349.

Hardwig, John. 1988. Evidence, Testimony, and the Problem of Individualism – A Response to Schmitt. Social Epistemology 2(4): 309‐321.

Hardwig, John. 1991. The Role of Trust in Knowledge. The Journal of Philosophy 88(12): 693‐708.

Hardwig, John. 1994. Toward an Ethics of Expertise. In Professional Ethics and Social Responsibility, ed. Daniel E. Wueste, 83‐101. London: Rowman and Littlefield.

Hedges, Larry V., Richard D. Laine & Rob Greenwald. 1994. Does Money Matter? A Meta‐ Analysis of Studies of the Effects of Differential School Inputs on Student Outcomes. Educational Researcher 23(3): 5‐14.

Hess, David J. 1997. Science Studies: An Advanced Introduction. New York: New York University Press.

Hilgartner, Stephen. 1990. The Dominant View of Popularization: Conceptual Problems, Political Uses. Social Studies of Science, 20(3): 519‐39.

Holmes, Lewis B. 1985. Response to Comments on “Teratogen Update: Bendectin”. Teratology 31(3): 432‐432.

Holton, Richard. 1994. Deciding to Trust, Coming to Believe. Australasian Journal of Philosophy 72(1): 63‐76.

Hopcroft, John E. et al. 2001. Introduction to Automata Theory, Languages, and Computation, 2nd ed. Boston: Addison‐Wesley.

Howson, Colin & Peter Urbach. 2006. Scientific Reasoning: The Bayesian Approach, 3rd ed. Chicago: Open Court.

Huber, Peter W. 1991. Galileo’s Revenge: Junk Science in the Courtroom. New York: Basic Books.

Hughes, Jeff. 2007. Insects or Neutrons? Science News Values in Interwar Britain. In Journalism, Science and Society: Science Communication between News and Public Relations, eds. Martin W. Bauer and Massimiano Bucchi, 11‐20. New York: Routledge.

‐ 225 ‐

Hume, David. 1748/1988. An Enquiry Concerning Human Understanding. Amherst, NY: Prometheus Books.

Intemann, Kristen. 2005. Feminism, Underdetermination, and Values in Science. Philosophy of Science 72(5): 1001–12.

Intemann, Kristen. 2009. Why Diversity Matters: Understanding and Applying the Diversity Component of the National Science Foundation’s Broader Impacts Criterion. Social Epistemology 23(3‐4): 249‐266.

Jasanoff, Sheila. 1995. Science at the Bar: Law, Science, and Technology in America . Cambridge, MA: Harvard University Press.

Jasanoff, Sheila. 1996. Is Science Socially Constructed – And Can It Still Inform Public Policy? Science and Engineering Ethics 2(3): 263‐276.

Jasanoff, Sheila. 2004. The Idiom of Co‐production. In States of Knowledge: The Co­ production of Science and Social Order, ed. Sheila Jasanoff, 1‐12. London: Routledge.

Junnarkar, Sandeep. 2002. Prime Efforts May Boost Encryption. CNET News.com (9 August) .

Kant, Immanuel. 1787/2007. Critique of Pure Reason, rev 2nd ed, trans. Norman Kemp Smith. New York: Palgrave Macmillan.

Keller, Evelyn Fox. 1983. A Feeling for the Organism: The Life and Work of Barbara McClintock. San Francisco: W.H. Freeman.

Kelly, Thomas. 2005. The Epistemic Significance of Disagreement . In Oxford Studies in Epistemology, ed. Tamar Gendler and John Hawthorne, 167‐196. New York: Oxford University Press.

Kingsley, Danny. 2002. A Prime Result. ABC Science Online (14 August) .

Kitcher, Philip. 1994. Contrasting Conceptions of Social Epistemology. In Socializing Epistemology: The Social Dimensions of Knowledge, ed. Frederick F. Schmitt, 111‐134. Lanham, MD: Rowman & Littlefield.

Kitcher, Philip. 2001. Science, Truth, and Democracy. New York: Oxford University Press.

Klayman, Joshua. 1995. Varieties of Confirmation Bias. Psychology of Learning and Motivation 32: 385‐418

Klein, Peter. 2009. Skepticism. In The Stanford Encyclopedia of Philosophy, Spring 2009 Edition, ed. Edward N. Zalta. http://plato.stanford.edu/archives/spr2009/entries/skepticism/.

‐ 226 ‐

Kourany, Janet. 2008. Replacing the Ideal of Value‐Free Science, In The Challenge of the Social and the Pressure of Practice: Science and Values Revisited, ed. Martin Carrier, Don Howard, and Janet Kourany, 87‐111. Pittsburgh, PA: University of Pittsburgh Press.

Kuhn, Thomas S. 1970. The Structure of Scientific Revolutions, 2nd ed. Chicago: The University of Chicago Press.

Kuhn, Thomas S. 1977. Objectivity, Value Judgment, and Theory Choice. In The Essential Tension: Selected Studies in Scientific Tradition and Change, 320‐39. Chicago: The University of Chicago Press.

Kunda, Ziva. 1990. The Case for Motivated Reasoning. Psychological Bulletin 108(3): 480‐498.

Kusch, Martin. 2002. Knowledge by Agreement: The Programme of Communitarian Epistemology. New York: Oxford University Press.

La Caze, Adam. 2009. Evidence‐Based Medicine Must Be. Journal of Medicine and Philosophy 34(5): 509‐27.

Lackey, Jennifer. 2007. Why We Don’t Deserve Credit for Everything We Know. Synthese 158(3): 345‐361.

Lackey, Jennifer. 2008. Learning from Words: Testimony as a Source of Knowledge. New York: Oxford University Press.

Lahno, Bernd. 2001. On the Emotional Character of Trust. Ethical Theory and Moral Practice 4: 171–189.

Lakatos, Imre. 1970. Falsification and the Methodology of Scientific Research Programmes. In Criticism and the Growth of Knowledge, eds. Imre Lakatos & Alan Musgrave, 91‐196. Cambridge, Cambridge University Press.

Latour, Bruno & Steve Woolgar. 1979. Laboratory Life: The Social Construction of Scientific Facts. Beverly Hills: Sage.

Latour, Bruno & Steve Woolgar. 1986. Laboratory Life: The Construction of Scientific Facts, 2nd ed. Princeton: Princeton University Press.

Latour, Bruno. 1987. Science in Action. Cambridge, MA: Harvard University Press.

Latour, Bruno. 2000. On The Partial Existence of Existing and Non‐Existing Objects. In Biographies of Scientific Objects, ed. Lorraine Daston, 247‐269. Chicago: University of Chicago Press.

Laudan, Larry. 1981. A Confutation of Convergent Realism. Philosophy of Science 48(1): 19‐49.

‐ 227 ‐

Laudan, Larry. 1984. Science and Values: The Aims of Science and Their Role in Scientific Debate. Berkeley: University of California Press.

Laudan, Larry. 1996. Beyond Positivism and Relativism: Theory, Method, and Evidence. Boulder, CO: Westview Press.

Lehrer, Keith & Carl Wagner. 1981. Rational Consensus in Science and Society: A Philosophical and Mathematical Study. Dordrecht: Reidel.

Lewenstein, Bruce V. 1995. From Fax to Facts: Communication in the Cold Fusion Saga. Social Studies of Science 25: 403‐436.

Lewis, David J. & Andrew Weigert. 1985. Trust as a Social Reality. Social Forces 63(4): 967‐985.

Lipton, Peter. 1998. The Epistemology of Testimony. Studies in History and Philosophy of Science 29(1): 1‐31.

Lipton, Peter. 2004. Inference to the Best Explanation, 2nd ed. London: Routledge.

Lipton, Peter. 2007. Alien Abduction: Inference to the Best Explanation and the Management of Testimony. Episteme: A Journal of Social Epistemology 4(3): 238‐51.

Longino, Helen & Ruth Doell. 1983. Body, Bias, and Behavior: A Comparative Analysis of Reasoning in Two Areas of Biological Science. Signs 9(2): 206‐227

Longino, Helen. 1990. Science as Social Knowledge: Values and Objectivity in Scientific Inquiry. Princeton: Princeton University Press.

Longino, Helen. 1994. The Fate of Knowledge in Social Theories of Science. In Socializing Epistemology: The Social Dimensions of Knowledge, ed. Frederick Schmitt, 135‐ 157. Lanham: Rowman and Littlefield.

Longino, Helen. 1995. Gender, Politics and Theoretical Virtues. Synthese 104: 383‐97.

Longino, Helen. 2002. The Fate of Knowledge. Princeton: Princeton University Press.

Longino, Helen. 2004. How Values Can Be Good for Science. In Science, Values and Objectivity, ed. Peter Machamer and Gereon Wolters, 127‐141. Pittsburgh, PA: University of Pittsburgh Press.

Longino, Helen. 2008. The Social Dimensions of Scientific Knowledge. In The Stanford Encyclopedia of Philosophy, Fall 2008 Edition, ed. Edward N. Zalta. http://plato.stanford.edu/archives/fall2008/entries/scientific‐knowledge‐ social/.

‐ 228 ‐

Machamer, Peter & Heather Douglas. 1999. Cognitive and Social Values. Science & Education 8: 45‐54.

MacKenzie, Donald A. 2001. Mechanizing Proof: Computing, Risk and Trust. Cambridge, MA: MIT Press.

Martin, Emily. 1991. The Egg and the Sperm: How Science Has Constructed a Romance Based on Stereotypical Male‐Female Roles. Signs 16(3): 485‐501.

McAllister, James W. 1997. Phenomena and Patterns in Data Sets. Erkenntnis 47: 217‐ 228.

Mercer, David. 2008. Science, Legitimacy, and “Folk Epistemology” in Medicine and Law: Parallels between Legal Reforms to the Admissibility of Expert Evidence and Evidence‐Based Medicine. Social Epistemology 22(4): 405‐423.

Merton, Robert K. 1973. The Sociology of Science: Theoretical and Empirical Investigations. Chicago: University of Chicago Press.

Michael, Mike. 1996. Ignoring Science: Discourses of Ignorance in the Public Understanding of Science. In Misunderstanding Science? The Public Reconstruction of Science and Technology, eds. lan Irwin & Brian Wynne, 107‐ 25. Cambridge: Cambridge University Press.

Mill, John S. 1859/1993. On Liberty. In Utilitarianism, On Liberty, Considerations on Representative Government, ed. J. M. Dent, 69‐187. London: Everyman.

Nelson, Lynn Hankinson. 1993. Epistemological Communities. In Feminist Epistemologies, eds. Linda Alcoff and Elizabeth Potter, 121‐159. New York: Routledge.

Nickerson, Raymond, S. 1998. Confirmation Bias: A Ubiquitous Phenomenon in Many Guises. Review of General Psychology 2(2): 175‐220.

Norton, John D. 2008. Must Evidence Underdetermine Theory? In The Challenge of the Social and the Pressure of Practice: Science and Values Revisited, eds. Martin Carrier, Don Howard & Janet Kourany, 189‐216. Pittsburgh, PA: University of Pittsburgh Press.

Nozick, Robert. 1981. Philosophical Explorations. Cambridge, MA: Harvard University Press.

Nye, Mary Jo. 1993. National Styles? French and English Chemistry in the Nineteenth and Early Twentieth Centuries. Osiris 8: 30‐49.

Odlyzko, Andrew. 2003. Alternatives to Peer Review I: Peer and Non‐Peer Review. In Peer Review in Health Sciences, 2nd ed., eds. Fiona Godlee & Tom Jefferson, 309‐ 11. London: BMJ Books.

‐ 229 ‐

Okruhlik, Kathleen. 1998. Gender and the Biological Sciences. In Philosophy of Science: The Central Issues, eds. M. Curd and J. A. Cover, 192‐208. New York: Norton.

Oxendine v. Merrell Dow Pharmaceuticals, Inc. 1996. WL 680992 (D.C. Super. 1996).

Pappas, George. 2005. Internalist vs. Externalist Conceptions of Epistemic Justification. The Stanford Encyclopedia of Philosophy, Fall 2008 Edition, ed. Edward N. Zalta. http://plato.stanford.edu/archives/fall2008/entries/justep‐intext/.

Parascandola, Mark. 1997. Chances, Individuals and Toxic Torts. Journal of Applied Philosophy 14(2): 147‐157.

Peirce, Charles S. 1877. The Fixation of Belief. Popular Science Monthly 12: 1‐15.

Pettit, Philip. 2006. No Testimonial Route to Consensus. Episteme: A Journal of Social Epistemology 3(3): 156‐165.

Phillips D. P. et al. 1991. Importance of the Lay Press in the Transmission of Medical Knowledge to the Scientific Community. The New England Journal of Medicine 325 (17 October): 1180‐83.

Pickering, Andrew. 1982. Interests and Analogies. In Science in Context: Reading in the Sociology of Science, eds. Barry Barnes and David Edges, 125‐146. Cambridge, MA: MIT press.

Pinch, Trevor. 1985. Towards an Analysis of Scientific Observation: The Externality and Evidential Significance of Observational Reports in Physics. Social Studies of Science 15(1): 3‐36.

Pinto, Meital. 2009. Who is Afraid of Language Rights in Israel? In The Multicultural Challenge in Israel, eds. Ohad Nachtomy & Avi Sagi, 26‐51. Boston: Academic Studies Press.

Plessy v. Ferguson. 1896. 163 U.S. 537.

Popper, Karl R. 1968. The Logic of Scientific Discovery, 3rd ed. London: Hutchinson.

Pradhan, Aharat. 2002. IIT Professor Makes Prime Mathematics Breakthrough. Rediff.com (10 August) .

Pritchard, Duncan. 2005. Epistemic Luck. New York: Oxford University Press.

Pritchard, Duncan. 2008. Sensitivity, Safety, and Anti‐Luck Epistemology. In The Oxford Handbook of Skepticism, ed. John Greco, 436‐455. New York: Oxford University Press.

‐ 230 ‐

Psillos, Stathis. 1994. A Philosophical Study of the Transition from the Caloric Theory of Heat to Thermodynamics: Resisting the Pessimistic Meta‐Induction. Studies in the History and Philosophy of Science 25(2): 159‐190.

Psillos, Stathis. 1999. Scientific Realism: How Science Tracks Truth. London: Routledge.

Quine, W. V. O. 1969. Epistemology Naturalized. In Ontological Relativity, and Other Essays, 69‐90. New York, Columbia University Press.

Rajghatta, Chidanand. 2002. India Still Has the Number on Maths. The Times of India (12 August) .

Ramachandran, R. 2002. A Prime Solution. Frontline: India’s National Magazine 19(17) (17 August) .

Raz, Joseph. 1984. Right‐Based Moralities. In Theories of Rights, ed. Jeremy Waldron, 182‐200. Oxford: Oxford University Press.

Réaume, Denise G. 1988. Individuals, Groups and Rights to Public Goods. The University of Toronto Law Journal 38(1): 1‐27.

Reid, Thomas. 1764/1997. An Inquiry into the Human Mind: On the Principles of Common Sense: A Critical Edition, ed. Derek R. Brookes. University Park, PA: Pennsylvania State University Press.

Rescher, Nicholas. 1993. Pluralism: Against the Demand for Consensus. New York: Oxford University Press.

Robinson, Sara. 2002. New Method Said to Solve Key Problem in Math. The New York Times (8 August): A20.

Rolin, Kristina. 2004. Why Gender Is a Relevant Factor in the Social Epistemology of Scientific Inquiry. Philosophy of Science 71(5): 880–891.

Rosen, Todd, Mary E. D'Alton, Lawrence D. Platt & Ronald Wapner. 2007. First‐ Trimester Ultrasound Assessment of the Nasal Bone to Screen for Aneuploidy. Obstetrics & Gynecology 110(2): 399‐404.

Sanders, Joseph. 1998. Benediction on Trial: A Study of Mass Tort Litigation. Ann Harbor: University of Michigan Press.

Schickore, Jutta. 2009. Studying Justificatory Practice: An Attempt to Integrate the History and Philosophy of Science. International Studies in the Philosophy of Science 23(1): 85‐107.

Schmitt, Frederick F. 1988. On the Road to Social Epistemic Interdependence. Social Epistemology 2(4): 297‐326.

‐ 231 ‐

Schmitt, Frederick F. 1994. The Justification of Group Beliefs. In Socializing Epistemology: The Social Dimensions of Knowledge, ed. Frederick F. Schmitt, 257‐87. Lanham, MD: Rowman & Littlefield.

Searle, John R. 1990/2002. Collective Intentions and Actions. In Consciousness and Language, 90‐105. Cambridge: Cambridge University Press .

Seavilleklein, Victoria. 2009. Challenging the Rhetoric of Choice in Prenatal Screening. Bioethics 23(1): 68‐77.

Shapin, Steven & Simon Schaffer. 1985. Leviathan and the Air­Pump: Hobbes, Boyle, and the Experimental Life. Princeton: Princeton University Press.

Shapin, Steven. 1994. A Social History of Truth: Civility and Science in Seventeenth­ Century England. Chicago: University of Chicago Press.

Sheppard, Celeste & Lawrence Platt. 2007. Nuchal Translucency and First Trimester Risk Assessment: A Systematic Review. Ultrasound Quarterly 23(2): 107‐116.

Silva, Michelle R. 2005. The Aerodynamics of Insects: The Role of Models and Matter in Scientific Experimentation. Social Epistemology 19(4): 325‐37.

Simon, Bart. 2001. Public Science: Media Configuration and Closure in the Cold Fusion Controversy. Public Understanding of Science 10(4): 383‐402.

Simon, Michael A. 1992. Causation, Liability and Toxic Risk Exposure. Journal of Applied Philosophy 9(1): 35‐44.

Sipser, Michael. 1997. Introduction to the Theory of Computation. Boston: PWS Pub.

Sismondo, Sergio. 1996. Science without Myth: On Constructions, Reality, and Social Knowledge. Albany, NY: State University of New York Press.

Sismondo, Sergio. 2010. An Introduction to Science and Technology Studies, 2nd ed. Oxford: Blackwell.

Solomon, Miriam & Alan Richardson. 2005. A Critical Context for Longino’s Critical Contextual Empiricism. Studies in History and Philosophy of Science 36: 211‐ 222.

Solomon, Miriam. 1994. Social Empiricism. Noûs 28(3): 325‐343.

Solomon, Miriam. 2001. Social Empiricism. Cambridge, MA: MIT Press.

Solomon, Miriam. 2007a. STS and Social Epistemology of Science. In The Handbook of Science and Technology Studies, 3rd ed., eds. Edward J. Hackett, Olga Amsterdamska, Michael Lynch, and Judy Wajcman, 241‐258. Cambridge, MA: MIT Press.

‐ 232 ‐

Solomon, Miriam. 2007b. The Social Epistemology of NIH Consensus Conferences. In Establishing Medical Reality: Essays in the Metaphysics and Epistemology of Biomedical Science, eds. Harold Kincaid and Jennifer McKitrick, 167‐177. Dordrecht: Springer.

Sommer, Marianne. 2006. Mirror, Mirror on the Wall: Neanderthal as Image and “Distortion” in Early 20th‐Century French Science and Press. Social Studies of Science 36(2): 207‐40.

Sosa, Ernest. 2007. Virtue Epistemology. Oxford: Oxford University Press.

Southern, David W. 1987. Gunnar Myrdal and Black­White Relations: The Use and Abuse of an American Dilemma, 1944­1969. Baton Rouge: Louisiana State University Press.

Stalnaker, Robert. 1984. Inquiry. Cambridge, MA: MIT Press.

Stanley, Jason. 2005. Knowledge and Practical Interests. Oxford: Oxford University Press.

Stegenga, Jacob. 2009. Robustness, Discordance, and Relevance. Philosophy of Science 76(5): 650‐661.

Steup, Matthias. 2008. The Analysis of Knowledge. In The Stanford Encyclopedia of Philosophy, Fall 2008 Edition, ed. Edward N. Zalta. http://plato.stanford.edu/archives/fall2008/entries/knowledge‐analysis/.

Stevens, Wesley M. 1985. Bede's Scientific Achievement. Newcastle: J & P Bealls.

Stiglic, Anton. 2005. The PRIMES is in P little FAQ. .

Strawson, Peter F. 1974. Freedom and Resentment. London: Methuen.

Tal, Eran. 2008. Do Scientific Instruments Have Objective Functions? Talk presented at Reclaiming the World: The Future of Objectivity, University of Toronto, Toronto, May 24.

Tassa, Tamir. 2002. With all Due Respect to the Deterministic Algorithm in Polynomial Time, the Code Will not Be Broken. Haaretz (26 August): 6. [Hebrew].

Taylor, Charles. 1992. Sources of the Self: The Making of Modern Identity. Cambridge, MA: Harvard University Press.

Thagard, Paul. 1992. Conceptual Revolutions. Princeton: Princeton University Press.

Thagard, Paul. 1997. Collaborative Knowledge. Noûs 31(2): 242‐261.

‐ 233 ‐

Thagard, Paul. 2000. Coherence in Thought and Action. Cambridge, MA: MIT Press.

Thagard, Paul. 2006. Testimony, Credibility, and Explanatory Coherence. Erkenntnis 63: 295–316.

Thompson, Clive. 2002. Outsider Math. The New York Times Magazine (15 December): 107.

Tucker, Aviezer. 2003. The Epistemic Significance of Consensus. Inquiry 46(4): 501‐21.

Tuomela, Raimo. 1992. Group Beliefs. Synthese 91(3): 285‐318.

Upshur, Ross. E. G. 2003. Are All Evidence‐Based Practices Alike? Problems in the Ranking of Evidence. Canadian Medical Association Journal 169(7): 672‐3.

Väliverronen, Esa. 1993. Science and the Media: Changing Relations. Science Studies 6(2): 23‐34. van der Sluijs, Josee, van Eijndhoven Jeroen, Simon Shackley & Brian Wynne. Anchoring Devices in Science for Policy: The Case of Consensus around Climate Sensitivity. Social Studies of Science 28(2): 291‐323. van Fraassen, Bas C. 1980. The Scientific Image. New York: Oxford University Press.

Vermeule, Adrian. 2009. Law and the Limits of Reason. New York: Oxford University Press.

Vitz, Rico. 2008. Doxastic Voluntarism. The Internet Encyclopedia of Philosophy, ed. James Fieser & Bradley Dowden. http://www.iep.utm.edu/doxa‐vol/

Vogel, Jonathan. 1990. Cartesian Skepticism and Inference to the Best Explanation. The Journal of Philosophy 87(11): 658‐666.

Wagner, Roy. 2009. Mathematical Marriages: Intercourse between Mathematics and Semiotic Choice. Social Studies of Science 39(2): 289‐308.

Webb, Mark. 2004. Can Epistemology Help? The Problem of the Kentucky‐Fried Rats. Social Epistemology 18(1): 51‐58.

Welbourne, Michael. 2001. Knowledge. Montreal: McGill‐Queen's University Press.

Whewell, William. 1858. Novum Organon Renovatum: Being the Second Part of the Philosophy of the Inductive Sciences. London: John W. Parker and Son.

Whitley, Richard. 1985. Knowledge Producers and Knowledge Acquirers: Popularisation as a Relation between Scientific Fields and Their Publics. In Expository Science: Forms and Functions of Popularisation, Sociology of the

‐ 234 ‐

Sciences Yearbook, vol. 9, eds. Terry Shinn & Richard Whitley, 3‐28. Dordrecht: Reidel.

Wilholt, Torsten. 2009. Bias and Values in Scientific Research. Studies in History and Philosophy of Science 40(1): 92‐101.

Wilson, Kumanan. 2010. Evidence‐Based Medicine. The Good the Bad and the Ugly. A Clinician's Perspective. Journal of Evaluation in Clinical Practice 16(2): 398‐400.

Woodward, James. 1989. Data and Phenomena. Synthese 79: 393‐472.

Worrall, John. 2002. What Evidence in Evidence‐Based Medicine? Philosophy of Science 69(3): S316‐S330.

Worrall, John. 2007a. Why There's No Cause to Randomize. The British Journal for the Philosophy of Science 58(3):451‐88.

Worrall, John. 2007b. Evidence in Medicine and Evidence‐Based Medicine. Philosophy Compass 2: 981‐1022.

Wray, K. Brad. 1999. A Defense of Longino's Social Epistemology. Philosophy of Science 66 (Supp.): S538‐S552.

Wray, K. Brad. 2001. Collective Belief and Acceptance. Synthese 129(3): 319‐333.

Wray, K. Brad. 2007. Who has Scientific Knowledge? Social Epistemology 21(3): 337‐ 347.

Wynne, Brian. 1996. Misunderstood Misunderstandings: Social Identities and the Public Uptake of Science. In Misunderstanding Science? The Public Reconstruction of Science and Technology, eds. Alan Irwin & Brian Wynne, 19‐46. Cambridge: Cambridge University Press.

‐ 235 ‐