The Economics of Artificial Intelligence

by Ajay Agrawal

  a period of very rapid substitution, with nonmarket- produced goods being

  substituted by market- produced goods. Women stopped making clothes

  and making pies and cakes from scratch, and started going to work, buying

  clothes, and buying pies and cake mixes. Technological change occurred in

  a way that crowded out homemade goods and crowded in women’s labor

  force participation.5 Should we think about this as increasing or decreasing

  work? One thing is clear, work shifted from outside our typical measure-

  ment scope to inside it. For example, I suspect that there are fewer childcare

  workers today than forty years ago if you count every stay- at- home mom

  with children as a childcare worker.

  Yet, time- use surveys reveal that the decline in hours worked is smaller

  than measured hours of employment suggest, at least since the 1970s. If we

  look at time- use surveys, dads are working more hours, even though they

  are working less in the labor force.6 Once we account for hours spent on

  childcare and housework, men work more hours than they did in the 1960s.

  Why consider childcare and housework hours? If we want to think about

  really measuring what happens to work we need a more holistic sense of

  5. Stevenson and Wolfers (2007).

  6. Council of Economic Advisors (2014).

  Artifi cial Intelligence, Income, Employment, and Meaning 195

  what work is. Particularly if the question is whether we can fi nd meaningful

  ways to spend our time outside of paid work.

  Artifi cial intelligence won’t replace the need for human connection, both

  in our personal lives and professionally. A robot may be able to care for an

  elderly bed- bound person, but it is unlikely to produce the joy and satisfac-

  tion of connecting with a human being. Will there be more paid jobs caring

  for one another? Undoubtedly. But will our higher incomes also allow us to

  choose to work less in order to provide more uncompensated care for our

  friends and family? I hope so.

  7.7 Productivity Growth Ultimately Gives Us

  Better Lives and More Options

  In the end, there are really two separate questions: there is an employment

  question, in which the fundamental question is, can we fi nd fulfi lling ways

  to spend our time if robots take our jobs? And there is an income question,

  can we fi nd a stable and fair distribution of income?

  The answer to both will depend on not just how technology changes, but

  how our institutions change in reaction to technological change. Do we

  embrace technology and increase funding for education, worker training,

  the arts, and community service? Or do we allow inequality to continue to

  grow unchecked, pitting workers against those investing in robots?

  The challenge for society is to ensure that we solve both problems. That

  we help shape a society in which people can fi nd fulfi lling ways to spend their

  time. And to solve that problem, we must also solve the separate problem of

  fi nding a stable and fair distribution of income.

  References

  Council of Economic Advisors. 2014. “Eleven Facts about American Families

  and Work.” https:// obamawhitehouse.archives .gov/ sites/ default/ fi les/ docs/ eleven _facts_about_family_and_work_fi nal .pdf.

  IGM Economic Experts Panel. 2014. Accessed Dec. 15, 2017. http:// www .igmchicago

  .org/ surveys/ robots.

  ———. 2017. Accessed Dec. 15, 2017. http:// www .igmchicago .org/ surveys/ robots

  - and- artifi cial- intelligence- 2.

  Organisation for Economic Co- operation and Development (OECD). 2017. “Hours

  Worked: Average Annual Hours Actually Worked.” OECD Employment and

  Labour Market Statistics (database). Accessed Sept. 13, 2017. https:// stats.oecd

  .org/ Index .aspx?DataSetCode=ANHRS.

  Stevenson, B., and J. Wolfers. 2007. “Marriage and Divorce: Changes and Their

  Driving Forces.” Journal of Economic Perspectives 21 (2): 27– 52.

  8

  Artifi cial Intelligence,

  Automation, and Work

  Daron Acemoglu and Pascual Restrepo

  8.1 Introduction

  The last two decades have witnessed major advances in artifi cial intel-

  ligence (AI) and robotics. Future progress is expected to be even more spec-

  tacular, and many commentators predict that these technologies will trans-

  form work around the world (Brynjolfsson and McAfee 2014; Ford 2016;

  Boston Consulting Group 2015; McKinsey Global Institute 2017). Recent

  surveys fi nd high levels of anxiety about automation and other technologi-

  cal trends, underscoring the widespread concerns about their eff ects (Pew

  Research Center 2017).

  These expectations and concerns notwithstanding, we are far from a sat-

  isfactory understanding of how automation in general, and AI and robotics

  in particular, impact the labor market and productivity. Even worse, much of

  the debate in both the popular press and academic circles centers around a

  false dichotomy. On the one side are the alarmist arguments that the oncom-

  ing advances in AI and robotics will spell the end of work by humans, while

  many economists on the other side claim that because technological break-

  throughs in the past have eventually increased the demand for labor and

  wages, there is no reason to be concerned that this time will be any diff erent.

  In this chapter, we build on Acemoglu and Restrepo (2016), as well as

  Daron Acemoglu is the Elizabeth and James Killian Professor of Economics at the Massa-

  chusetts Institute of Technology and a research associate of the National Bureau of Economic Research. Pascual Restrepo is assistant professor of economics at Boston University.

  We are grateful to David Autor for useful comments. We gratefully acknowledge fi nancial support from Toulouse Network on Information Technology, Google, Microsoft, IBM, and

  the Sloan Foundation. For acknowledgments, sources of research support, and disclosure of the authors’ material fi nancial relationships, if any, please see http:// www .nber .org/ chapter

  / c14027.ack.

  197

  198 Daron Acemoglu and Pascual Restrepo

  Zeira (1998) and Acemoglu and Autor (2011) to develop a framework for

  thinking about automation and its impact on tasks, productivity, and work.

  At the heart of our framework is the idea that automation and thus AI

  and robotics replace workers in tasks that they previously performed, and

  via this channel, create a powerful displacement eff ect. In contrast to pre-

  sumptions in much of macroeconomics and labor economics, which main-

  tain that productivity- enhancing technologies always increase overall labor

  demand, the displacement eff ect can reduce the demand for labor, wages,

  and employment. Moreover, the displacement eff ect implies that increases

  in output per worker arising from automation will not result in a propor-

  tional expansion of the demand for labor. The displacement eff ect causes

  a decoupling of wages and output per worker, and a decline in the share of

  labor in national income.

  We then highlight several countervailing forces that push against the

  displacement eff ect and may imply that automation, AI, and robotics could

  increase labor demand. First, the substitution of cheap machines for human

  labor c
reates a productivity eff ect: as the cost of producing automated tasks declines, the economy will expand and increase the demand for labor in

  nonautomated tasks. The productivity eff ect could manifest itself as an

  increase in the demand for labor in the same sectors undergoing automa-

  tion or as an increase in the demand for labor in nonautomating sectors.

  Second, capital accumulation triggered by increased automation (which

  raises the demand for capital) will also raise the demand for labor. Third,

  automation does not just operate at the extensive margin—replacing tasks

  previously performed by labor—but at the intensive margin as well, increas-

  ing the productivity of machines in tasks that were previously automated.

  This phenomenon, which we refer to as deepening of automation, creates a

  productivity eff ect but no displacement, and thus increases labor demand.

  Though these countervailing eff ects are important, they are generally

  insuffi

  cient to engender a “balanced growth path,” meaning that even if

  these eff ects were powerful, ongoing automation would still reduce the share

  of labor in national income (and possibly employment). We argue that there

  is a more powerful countervailing force that increases the demand for labor

  as well as the share of labor in national income: the creation of new tasks,

  functions and activities in which labor has a comparative advantage rela-

  tive to machines. The creation of new tasks generates a reinstatement eff ect

  directly counterbalancing the displacement eff ect.

  Indeed, throughout history we have not just witnessed pervasive automa-

  tion, but a continuous process of new tasks creating employment opportuni-

  ties for labor. As tasks in textiles, metals, agriculture, and other industries

  were being automated in the nineteenth and twentieth centuries, a new range

  of tasks in factory work, engineering, repair, back- offi

  ce, management, and

  fi nance generated demand for displaced workers. The creation of new tasks

  Artifi cial Intelligence, Automation, and Work 199

  is not an autonomous process advancing at a predetermined rate, but one

  whose speed and nature are shaped by the decisions of fi rms, workers, and

  other actors in society, and might be fueled by new automation technologies.

  First, this is because automation, by displacing workers, may create a greater

  pool of labor that could be employed in new tasks. Second, the currently

  most discussed automation technology, AI itself, can serve as a platform to

  create new tasks in many service industries.

  Our framework also highlights that even with these countervailing forces,

  the adjustment of an economy to the rapid rollout of automation tech-

  nologies could be slow and painful. There are some obvious reasons for

  this related to the general slow adjustment of the labor market to shocks,

  for example, because of the costly process of workers being reallocated to

  new sectors and tasks. Such reallocation will involve both a slow process

  of searching for the right matches between workers and jobs, and also the

  need for retraining, at least for some of the workers.

  A more critical, and in this context more novel, factor is a potential mis-

  match between technology and skills—between the requirements of new

  technologies and tasks and the skills of the workforce. We show that such

  a mismatch slows down the adjustment of labor demand, contributes to

  inequality, and also reduces the productivity gains from both automation

  and the introduction of new tasks (because it makes the complementary

  skills necessary for the operation of new tasks and technologies more scarce).

  Yet another major factor to be taken into account is the possibility of

  excessive automation. We highlight that a variety of factors (ranging from a

  bias in favor of capital in the tax code to labor market imperfections create

  a wedge between the wage and the opportunity cost of labor) and will push

  toward socially excessive automation, which not only generates a direct inef-

  fi ciency, but also acts as a drag on productivity growth. Excessive automa-

  tion could potentially explain why, despite the enthusiastic adoption of new

  robotics and AI technologies, productivity growth has been disappointing

  over the last several decades.

  Our framework underscores as well that the singular focus of the research

  and the corporate community on automation, at the expense of other types

  of technologies including the creation of new tasks, could be another factor

  leading to a productivity slowdown because it forgoes potentially valuable

  productivity growth opportunities in other domains.

  In the next section, we provide an overview of our approach without

  presenting a formal analysis. Section 8.3 introduces our formal framework,

  though to increase readability, our presentation is still fairly nontechnical

  (and formal details and derivations are relegated to the appendix). Section

  8.4 contains our main results, highlighting both the displacement eff ect

  and the countervailing forces in our framework. Section 8.5 discusses the

  mismatch between skills and technologies, potential causes for slow pro-

  200 Daron Acemoglu and Pascual Restrepo

  ductivity growth and excessive automation, and other constraints on labor

  market adjustment to automation technologies. Section 8.6 concludes, and

  the appendix contains derivations and proofs omitted from the text.

  8.2 Automation, Work, and Wages: An Overview

  At the heart of our framework is the observation that robotics and current

  practice in AI are continuing what other automation technologies have done

  in the past: using machines and computers to substitute for human labor in

  a widening range of tasks and industrial processes.

  Production in most industries requires the simultaneous completion of

  a range of tasks. For example, textile production requires production of

  fi ber, production of yarn from fi ber (e.g., by spinning), production of the

  relevant fabric from the yarn (e.g., by weaving or knitting), pretreatment

  (e.g., cleaning of the fabric, scouring, mercerizing and bleaching), dyeing

  and printing, fi nishing, as well as various auxiliary tasks including design,

  planning, marketing, transport, and retail.1 Each one of these tasks can be

  performed by a combination of human labor and machines. At the dawn

  of the British Industrial Revolution, most of these tasks were heavily labor

  intensive. Many of the early innovations of that era were aimed at automat-

  ing spinning and weaving by substituting mechanized processes for the labor

  of skilled artisans (Mantoux 1928).2

  The mechanization of US agriculture off ers another example of machines

  replacing workers in tasks they previously performed (Rasmussen 1982).

  In the fi rst half of the nineteenth century, the cotton gin automated the

  labor- intensive process of separating the lint from the cotton seeds. In the

  second half of the nineteenth century, horse- powered reapers, harvesters,

  and plows replaced manual labor working with more rudimentary tools such

  as hoes, sickles, and scythes, and this process was continued with tractors

  in the twenti
eth century. Horse- powered threshing machines and fanning

  mills replaced workers employed in threshing and winnowing, two of the

  most labor- intensive tasks left in agriculture at the time. In the twentieth

  century, combine harvesters and a variety of other mechanical harvest-

  ers improved upon the horse- powered machinery, and allowed farmers to

  mechanically harvest several diff erent crops.

  Yet another example of automation comes from the development of the

  1. See http:// textileguide .chemsec .org/ fi nd/ get- familiar- with- your- textile- production

  - processes/.

  2. It was this displacement eff ect that motivated Luddites to smash textile machines and agricultural workers during the Captain Swing riots to destroy threshing machines. Though these workers often appear in history books as misguided, there was nothing misguided about their economic fears. They were quite right that they were going to be displaced. Of course, had they been successful, they might have prevented the Industrial Revolution from gaining momentum with potentially disastrous consequences for technological development and our subsequent prosperity.

  Artifi cial Intelligence, Automation, and Work 201

  factory system in manufacturing and its subsequent evolution. Beginning

  in the second half of the eighteenth century, the factory system introduced

  the use of machine tools such as lathes and milling machines, replacing

  the more labor- intensive production techniques relying on skilled artisans

  (Mokyr 1990). Steam power and later electricity greatly increased the oppor-

  tunities for the substitution of capital for human labor. Another important

  turning point in the process of factory automation was the introduction

  of machines controlled via punch cards and then numerically controlled

  machines in the 1940s. Because numerically controlled machines were more

  precise, faster, and easier to operate than manual technologies, they enabled

  signifi cant cost savings while also reducing the role of craft workers in manu-

  facturing production. This process culminated in the widespread use of

  CNC (computer numerical control) machinery, which replaced the numeri-

  cally controlled vintages (Groover 1983). A major new development was the

  introduction of industrial robots in the late 1980s, which automated many

  of the remaining labor- intensive tasks in manufacturing, including machin-