WHAT JOBS ARE AFFECTED BY AI? Better-paid, better-educated workers face the most exposure MARK MURO, JACOB WHITON, and ROBERT MAXIM November 2019 Contents Introduction 3 What we know about AI, and what we don’t 5 Approach: Using AI to assess AI’s workforce impacts 8 Findings 11 Discussion 22 References 25 Endnotes 26 Appendix 28 Acknowledgements 44 WHAT JOBS ARE AFFECTED BY AI? 2 Introduction he debate between experts over how better-educated workers will largely make out Tautomation will affect the future of work has alright as automation spreads. This result was not been one of the most active cottage industries an outlier. Similar findings have accumulated in in labor economics in recent years. Numerous numerous reports, ranging from those by teams scholars forecast major disruptions of human at Oxford University and the OECD to the African work; others minimize those impacts. American Mayors Association. And yet, the field has nevertheless managed to But what about artificial intelligence (AI), the generate a number of shared insights, with none increasingly powerful form of digital automation more consistent than the finding that least well- using machines that can learn, reason, and act off will suffer automation’s greatest shocks on for themselves? the labor market. In recent years, AI applications have generated “The vulnerable will be the most vulnerable” increasing interest in “future of work” discussions was a key takeaway of the report on AI and as the technology achieved superhuman automation we released earlier this year, for performance in a range of valuable tasks, from example. That analysis, based on forecasts radiology to legal contracts. However, it has been of occupation-level automation exposure difficult to get a specific read on AI’s implications from expert assessment by the McKinsey on the labor market. Global Institute, showed that higher-wage, WHAT JOBS ARE AFFECTED BY AI? 3 In part because the technologies have not fully on statistical associations, as opposed to yet been widely adopted, analyses such as relying in large part on expert prognostications. Brookings’s or those from Oxford, OECD, and McKinsey have had to rely either on case What do we find in working with Webb’s studies or subjective assessments by experts data? Above all, that Webb’s AI measures depict a to determine which occupations might be very different range of impacts on the workforce susceptible to an AI takeover. What’s more, none than those from robotics and software. Where of these analyses focused solely and specifically the robotics and software that dominate the on AI. Instead, most research has concentrated automation field seem mostly to involve “routine” on an undifferentiated array of “automation” or “rule-based” tasks (and thus lower- or middle- technologies including robotics, software, and pay roles), AI’s distinctive capacities suggest that AI all at once. The result has been a lot of higher-wage occupations will be some of the discussion—but not a lot of clarity—about AI, with most exposed. prognostications that range from utopian to apocalyptic.1 Unlike robotics (associated with the factory floor) and computers (associated with routine But now comes a new approach. By quantifying office activities), AI has a distinctly white-collar the overlap between the text of AI patents and bent. While earlier waves of automation have led the text of job descriptions, Stanford University to disruption across the lower half of the wage Ph.D. candidate Michael Webb has developed an distribution, AI appears likely to have different elegant new way to identify the kinds of tasks and impacts, with its own windfalls and challenges. occupations likely to be affected by particular White-collar, well-paid America—radiologists, legal AI capabilities—and has graciously shared his professionals, optometrists, and many more—will “exposure scores” for occupations to allow likely get no free pass on this flavor of digital further analysis by Brookings.2 In doing so, Webb disruption. has allowed us to further test a new analytic approach that is extremely important, as it allows Given the potential of these technologies, it us to probe the kinds of occupations likely to be behooves us to get a clearer read on their labor affected by AI specifically, as opposed to those market reach, which is what the following pages affected by the broader swath of automation begin to do. technologies. With these data we are able to rely WHAT JOBS ARE AFFECTED BY AI? 4 What we know about AI, and what we don’t irst, some context: What is AI, and why are that “intelligence” has always been defined as Fits workforce impacts so hard to assess? This whatever it is that humans can do that computers is an important question, because the problem cannot. But since that frontier has been changing of gauging its effects owes to the disparate, rapidly, the definition doesn’t limit the field much. changing nature of AI itself, which draws on an ever-evolving set of algorithms and approaches The definitional problem does not disappear even to generating machines with human-level if the aperture is narrowed to focus on machine intelligence. learning (ML)—the branch of statistics on which most AI currently depends. Machine learning What is AI? can be straightforwardly defined as computers’ use of algorithms to find statistical patterns in Part of the challenge of analyzing AI in massive amounts of data, which can then be used general is that no single definition of the to make predictions. Such statistical pattern- technology serves to pin down its operations finding has been around for decades, but this and capabilities.3 It’s only somewhat helpful to field also is evolving rapidly. Recent years have say that AI involves programming computers seen a surge in improved algorithms that have to do things which—if done by humans—would been accelerated by advances in computer speed, be said to require “intelligence,” whether it be data collection, and storage, driving an explosion planning, learning, reasoning, problem-solving, of improved applications including image perception, or prediction.4 The problem here is recognition, voice interpretation, preference WHAT JOBS ARE AFFECTED BY AI? 5 prediction, autonomy, and decision support.5 This well-established, well-understood robotics and explosion of applications is continually changing software technologies (several types of AI were the nature and boundaries of ML, adding to the included too). difficulty of defining and analyzing AI as a field and set of applications. AI presents a more challenging set of issues. Even by reducing the scope of the present analysis to Why AI’s workplace impacts are ML applications, analysis of AI on the workforce must contend with a profusion of relatively new, hard to assess hard-to-discern technologies that have not yet been widely adopted by firms or diffused far Because AI is such a moving target, efforts to across the economy into practical use. assess its impacts on work are also complicated. Our earlier report on automation explained how Contrary to robotics and software, for example, assessments of well-recognized technologies— researchers have had little time to learn about heavily dependent on experts’ experience—could AI’s primary use cases in the economy—as is be employed to identify the aspects of jobs most indicated by Figure 1, which tracks the recent susceptible to particular technologies. In that emergence of machine learning patenting.6 instance, our analyses were made easier by the availability of McKinsey’s expert forecasts, which reflected extensive experience with relatively Figure 1. Index of patent counts by technology Patent counts by technology, 1980 - 2016 SEMICONDUCTORS SOFTWARE MACHINE LEARNING 30,000 60,000 600 25,000 50,000 500 20,000 40,000 400 15,000 30,000 300 10,000 20,000 200 5,000 10,000 100 0 0 0 2 2 2 2 2 2 6 6 6 6 6 6 4 4 4 4 4 4 8 8 8 8 8 8 0 0 0 0 0 0 1 1 1 1 1 1 9 9 9 9 9 9 8 0 8 0 8 0 8 0 8 0 8 0 8 8 8 0 0 0 0 0 0 0 0 0 9 9 9 9 9 9 9 9 9 0 0 0 9 9 9 0 0 0 9 9 9 0 0 0 2 2 2 2 2 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 Source: Webb, Short, Bloom, and Lerner (2018) “Some facts of high-tech patenting.” WHAT JOBS ARE AFFECTED BY AI? 6 Consequently, as the scholars Erik Brynjolfsson AI-exposed jobs) using multicriteria subjective and Tom Mitchell have written, there is “no widely rubrics informed by their deep knowledge of the shared agreement on the tasks where machine field.8 However, even they express humility about learning systems excel, and thus little agreement such efforts. The general takeaway is that the on the expected impacts on the workforce and on evolving, emergent nature of AI poses a tough the economy more broadly.”7 challenge for analyses of its impact—especially those that rely on standard expert judgement. Brynjolfsson and Mitchell have done their best Another approach is needed. to identify AI-suitable tasks (and therefore WHAT JOBS ARE AFFECTED BY AI? 7 Approach: Using AI to assess AI’s workforce impacts his is where Michael Webb’s new approach these technologies. (For more on the process Tcomes in. To circumvent many of the problems see Michael Webb, “The Impact of Artificial posed by AI for labor market analysis, this brief Intelligence on the Labor Market” and its uses the outputs of Webb’s novel AI method for appendices.) quantifying the “exposure” of occupations to assess the broader labor market impacts.