Cambridge Journal of Regions, Economy and Society 00, 00, 0–0 doi:10.1093/cjres/rsz022 Downloaded from https://academic.oup.com/cjres/advance-article-abstract/doi/10.1093/cjres/rsz022/5680462 by MIT Libraries user on 19 December 2019 The wrong kind of AI? Artificial intelligence and the future of labour demand Daron Acemoglua and Pascual Restrepob aMIT, Cambridge, MA, USA, [email protected] bBoston University, Boston, MA, USA, [email protected] Received on May 30, 2019; editorial decision on October 16, 2019; accepted on November 18, 2019 Artificial intelligence (AI) is set to influence every aspect of our lives, not least the way pro- duction is organised. AI, as a technology platform, can automate tasks previously performed by labour or create new tasks and activities in which humans can be productively employed. Recent technological change has been biased towards automation, with insufficient focus on creating new tasks where labour can be productively employed. The consequences of this choice have been stagnating labour demand, declining labour share in national income, rising inequality and lowering productivity growth. The current tendency is to develop AI in the direction of further automation, but this might mean missing out on the promise of the ‘right’ kind of AI, with better economic and social outcomes. Keywords: automation, artificial intelligence, jobs, inequality, innovation, labour demand, productivity, tasks, technology, wages JEL Classifications:J23, J24 Artificial intelligence (AI) is one of the most we can still shape the direction of AI research promising technologies currently being de- and the future of work. veloped and deployed. Broadly speaking, AI refers to the study and development of “intel- ligent (machine) agents”, which are machines, AI as a technology platform softwares or algorithms that act intelligently Human (or natural) intelligence comprises sev- by recognising and responding to their envir- eral different types of mental activities. These onment.1 There is a lot of excitement, some include simple computation, data processing, hype and a fair bit of apprehension about pattern recognition, prediction, various types of what AI will mean for our security, society problem solving, judgment, creativity, and com- and economy. But a critical question has been munication. Early AI, pioneered in the 1950s by largely overlooked: are we investing in the researchers from computer science, psychology “right” type of AI, the kind with the greatest po- and economics, such as Marvin Minsky, Seymour tential for raising productivity and generating Papert, John McCarthy, Herbert Simon and broad-based prosperity? We do not have a de- Allen Newell, sought to develop machine intel- finitive answer right now—nobody does. But ligence capable of performing all of these dif- this is the right time to ask this question while ferent types of mental activities.2 The goal was © The Author(s) 2019. Published by Oxford University Press on behalf of the Cambridge Political Economy Society. All rights reserved. For permissions, please email: [email protected] Acemoglu and Restrepo Downloaded from https://academic.oup.com/cjres/advance-article-abstract/doi/10.1093/cjres/rsz022/5680462 by MIT Libraries user on 19 December 2019 nothing short of creating truly intelligent ma- Even if we focus on its narrow version, AI chines. Herbert Simon and Allen Newell, for should be thought of as a technology plat- example, claimed in 1958 “there are now in the form—there are many ways AI technology can world machines that think, that learn and that be developed as a commercial or production create. Moreover, their ability to do these things technology, with widely varying applications. is going to increase rapidly until—in a visible This matters greatly because it implies that the future—the range of problems they can handle economic and social consequences of AI tech- will be coextensive with the range to which the nologies are not preordained but depend on human mind has been applied.”3 how we decide to advance and build on this These ambitious goals were soon dashed. AI platform. To some degree, this is true of all clus- came back into fashion in the 1990s, but with ters of technologies, but it is more emphatically a different and more modest ambition: to rep- so for AI.4 To see this, contrast it with a related licate and then improve upon human intelli- but distinct new technology, robotics. Robotics gence in pattern recognition and prediction often makes use of AI and other digital tech- (pre-AI computers were already better than nologies for processing data, but is distin- humans in computation and data processing). guished from other digital technologies by its Many decision problems and activities we rou- focus on interacting with the physical world tinely engage in can be viewed as examples (moving around, transforming, rearranging or of pattern recognition and prediction. These joining objects). Industrial robots are already include recognising faces (from visual data), widespread in many manufacturing industries recognising speech (from auditory data), rec- and in some retail and wholesale establish- ognising abstract patterns in data we are pre- ments. But their economic use is quite specific, sented with, and making decisions on the and centres on automation of narrow tasks, that basis of past experience and current informa- is, substituting machines for certain specific ac- tion. Though there are researchers working on tivities and functions previously performed by “Artificial General Intelligence”, much of the humans.5 research and almost all commercial applica- tions of AI are in these more modest domains referred to as “Narrow AI”—even if the rele- Implications of technology for work vant applications are numerous and varied. The and labour big breakthroughs and the renewed excitement How do new technologies impact the nature of in AI are coming from advances in hardware production and work? Employment and wages and algorithms that enable the processing and of different types of workers? The standard analysis of vast amounts of unstructured data approach, both in popular discussions and in (for example, speech data that cannot be repre- academic writings, presumes that any advance sented in the usual structured ways, such as in that increases productivity (value added per simple, Excel-like databases). Central to this re- worker) also tends to raise the demand for naissance of AI have been methods of machine labour, and thus employment and wages. Of learning (which are the statistical techniques course, technological progress might lead to that enable computers and algorithms to learn, job loss in some sectors. But even when that predict and perform tasks from large amounts happens, the standard narrative goes, other sec- of data without being explicitly programmed) tors will expand and contribute to overall em- and what is called “deep learning” (algorithms ployment and wage growth. Moreover, even if that use multi-layered programs, such as neural technological progress benefits some workers nets, for improved machine learning, statistical more than others and increases inequality, the inference and optimisation). standard approach still predicts that it will Page 2 of 11 The wrong kind of AI? Downloaded from https://academic.oup.com/cjres/advance-article-abstract/doi/10.1093/cjres/rsz022/5680462 by MIT Libraries user on 19 December 2019 tend to raise the labour demand for all types not reduce overall labour demand. A second of workers.6 conclusion is that whether they reduce overall This view is critically underpinned by the labour demand depends on the strength of the way in which the economic impact of new tech- productivity effect. nology is conceptualised—as enabling labour This last observation has important implica- to become more productive in pretty much all tions: contrary to popular claims that the future of the activities and tasks that it performs. Yet, of labour is threatened by “brilliant” new tech- this not only lacks descriptive realism (what nologies, the greater danger for labour comes technology makes labour uniformly more pro- from technology that is not raising productivity ductive in everything?), but may paint an ex- sufficiently. In particular, if new automation cessively rosy picture of the implications of new technologies are not great but just “so-so” (just technologies. Indeed, in such a world Luddites’ good enough to be adopted but not so much concerns about the disruptive and job displacing more productive than the labour they are re- implications of technology would be misplaced, placing), there is a double jeopardy for labour— and they would have smashed all of those ma- there is a displacement effect, taking passed chines in vain. away from labour, but no powerful productivity The reality of technological change is rather gains redressing some of the decline in labour different. Many new technologies—those we demand generated by the displacement effects. call automation technologies—are not intended Is this far-fetched? Not really. We have previ- to increase labour’s productivity, but are ex- ously studied the implications of one of the most plicitly aimed at replacing it by substituting important automation technologies, industrial cheaper capital (machines) in a range of tasks robots.9 Industrial robots are not technologies performed by humans.7 As a result, automa- aimed at increasing labour’s productivity but tion technologies, by displacing workers from are designed to automate tasks that were previ- the tasks they were previously performing, al- ously performed by production workers on the ways reduce the labour’s share in value added. factory floor. The evidence is fairly clear that Put differently, these technologies raise prod- industries where more industrial robots are uctivity by more than wages and employment. introduced experience declines in labour de- They may even reduce overall labour demand mand (especially for production workers) and (and thus reduce wages, employment or both). sizable falls in their labour share.