Artificial intelligence: How does it work, why does it matter, and what can we do about it? STUDY Panel for the Future of Science and Technology EPRS | European Parliamentary Research Service Author: Philip Boucher Scientific Foresight Unit (STOA) PE 641.547 – June 2020 EN Artificial intelligence: How does it work, why does it matter, and what can we do about it? Artificial intelligence (AI) is probably the defining technology of the last decade, and perhaps also the next. The aim of this study is to support meaningful reflection and productive debate about AI by providing accessible information about the full range of current and speculative techniques and their associated impacts, and setting out a wide range of regulatory, technological and societal measures that could be mobilised in response. AUTHOR Philip Boucher, Scientific Foresight Unit (STOA), This study has been drawn up by the Scientific Foresight Unit (STOA), within the Directorate-General for Parliamentary Research Services (EPRS) of the Secretariat of the European Parliament. To contact the publisher, please e-mail [email protected] LINGUISTIC VERSION Original: EN Manuscript completed in June 2020. DISCLAIMER AND COPYRIGHT This document is prepared for, and addressed to, the Members and staff of the European Parliament as background material to assist them in their parliamentary work. The content of the document is the sole responsibility of its author(s) and any opinions expressed herein should not be taken to represent an official position of the Parliament. Reproduction and translation for non-commercial purposes are authorised, provided the source is acknowledged and the European Parliament is given prior notice and sent a copy. Brussels © European Union, 2020. PE 641.547 ISBN: 978-92-846-6770-3 doi: 10.2861/44572 QA-01-20-338-EN-N http://www.europarl.europa.eu/stoa (STOA website) http://www.eprs.ep.parl.union.eu (intranet) http://www.europarl.europa.eu/thinktank (internet) http://epthinktank.eu (blog) II Artificial intelligence: How does it work, why does it matter, and what can we do about it? Executive summary Artificial intelligence (AI) is probably the defining technology of the last decade, and perhaps also the next. The aim of this study is to support meaningful reflection and productive debate about AI by providing accessible information about the full range of current and speculative techniques and their associated impacts, and setting out a wide range of regulatory, technological and societal measures that could be mobilised in response. What is artificial intelligence? The study adopts the European Commission's 2018 definition of AI, which is both accessible and typical of contemporary definitions. AI refers to systems that display intelligent behaviour by analysing their environment and taking action – with some degree of autonomy – to achieve specific goals. Since AI refers to so many techniques and contexts, greater precision is required in order to hold meaningful and constructive debates about it. For example, arguments about simple 'expert systems' used in advisory roles need to be distinguished those from those concerning complex data- driven algorithms that automatically implement decisions about individuals. Similarly, it is important to distinguish arguments about speculative future developments that may never occur from those about current AI that already affects society today. How does artificial intelligence work? Chapter 2 sets out accessible introductions to some of the key techniques that come under the AI banner. They are grouped into three sections, which gives a sense of the chronology of the development of different approaches. The first wave of early AI techniques is known as 'symbolic AI' or expert systems. Here, human experts create precise rule-based procedures – known as 'algorithms' – that a computer can follow, step by step, to decide how to respond intelligently to a given situation. Fuzzy logic is a variant of the approach that allows for different levels of confidence about a situation, which is useful for capturing intuitive knowledge, so that the algorithm can make good decisions in the face of wide-ranging and uncertain variables that interact with each other. Symbolic AI is at its best in constrained environments which do not change much over time, where the rules are strict and the variables are unambiguous and quantifiable. While these methods can appear dated, they remain very relevant and are still successfully applied in several domains, earning the endearing nickname 'good old- fashioned AI'. The second wave of AI comprises more recent 'data-driven' approaches which have developed rapidly over the last two decades and are largely responsible for the current AI resurgence. These automate the learning process of algorithms, bypassing the human experts of first wave AI. Artificial neural networks (ANNs) are inspired by the functionality of the brain. Inputs are translated into signals which are passed through a network of artificial neurons to generate outputs that are interpreted as responses to the inputs. Adding more neurons and layers allow ANNs to tackle more complex problems. Deep learning simply refers to ANNs with several layers. Machine learning (ML) refers to the transformation of the network so that these outputs are considered useful – or intelligent – responses to the inputs. ML algorithms can automate this learning process by making gradual improvements to individual ANNs, or by applying evolutionary principles to yield gradual improvements in large populations of ANNs. III STOA | Panel for the Future of Science and Technology The third wave of AI refers to speculative possible future waves of AI. While first and second wave techniques are described as 'weak' or 'narrow' AI in the sense that they can behave intelligently in specific tasks, 'strong' or 'general' AI refers to algorithms that can exhibit intelligence in a wide range of contexts and problem spaces. Such artificial general intelligence (AGI) is not possible with current technology and would require paradigm shifting advancement. Some potential approaches have been considered, including advanced evolutionary methods, quantum computing and brain emulation. Other forms of speculative future AI such as self-explanatory and contextual AI can seem modest in their ambitions, but their potential impact – and barriers to implementation – should not be underestimated. Why does artificial intelligence matter? Chapter 3 builds upon the understanding of how these technologies work to examine several opportunities and challenges presented by their application in various contexts. Several challenges are associated with today's AI. Broadly, they can be understood as a balancing act between avoiding underuse whereby we miss out on potential opportunities, and avoiding overuse whereby AI is applied for tasks for which it is not well suited or results in problematic outcomes. The ML process makes some algorithms vulnerable to bias, and their complexity makes their decision- making logic difficult to understand and explain. There are some important challenges in ensuring that the costs and benefits of AI development are distributed evenly, avoiding the concentration of resources in uncompetitive markets and prioritising applications that alleviate rather than exacerbate existing structural inequalities. Other key challenges include the public acceptability of the technology, its alignment with social values, and concerns about some military applications. There are also several longer-term opportunities and challenges that are contingent upon future developments which might never happen. Some utopian and dystopian scenarios might contribute to hype cycles, but they also present an opportunity to prepare for more moderate trends and reflect upon what we want from the technology. For example, it has been suggested that AI could lead to major job losses or make the concept of employment obsolete, that it could escape human control and take control of its own development, that it could challenge human autonomy or develop artificial emotions or consciousness, presenting interesting – yet speculative – philosophical questions. What can we do about artificial intelligence? Chapter 4 sets out several options that could be mobilised in response to the opportunities and challenges that were set out in the previous chapter. The options are organised into three sections, focussing on policy, technology and society. Each section contains seven themes with several options for each, with over 100 measures in total. Most AI policy debates concern how to shape the regulatory and economic context in which AI is developed and applied in order to respond to specific opportunities and challenges. These could include creating a supportive economic and policy context, promoting more competitive ecosystems, improving the distribution of benefits and risks, building resilience against a range of problematic outcomes, enhancing transparency and accountability, ensuring mechanisms for liability and developing governance capacity. There are also more abstract policy debates about the broad regulatory approach. This includes questions about whether to have regulation that specifically targets AI, or to regulate it by applying and updating tech-neutral mechanisms that apply to all activities, regardless of whether they use AI. IV Artificial intelligence: How does it work, why does it matter, and what can we do about it? Similarly, there are institutional debates about whether