The Economics of Artificial Intelligence Read online

  technical advance. Furthermore, economists possess a vast methodological

  arsenal that may prove very useful for that purpose—we should not shy away

  from stepping into this area, since its importance for the economy cannot

  be overstated.” The next set of chapters also emphasize the distributional

  challenges of economic growth driven by rapid technological change.

  Growth, Jobs, and Inequality

  Much of the popular discussion around AI focuses on the impact on jobs.

  If machines can do what humans do, then will there still be work for humans

  in the future? The chapters in this section dig into the consequences of AI

  for jobs, economic growth, and inequality. Almost all chapters emphasize

  that technological change means an increase in wealth for society. As Jason

  Furman puts it in chapter 12, “We need more artifi cial intelligence.” At the

  same time, it is clear that the impact of AI on society will depend on how

  the increased income from AI is distributed. The most recent GPTs to dif-

  fuse, computers and the internet, likely led to increased inequality due to

  skill- bias (e.g., Autor, Katz, and Krueger 1998; Akerman, Gaarder, and

  Mogstad 2015) and to an increased capital share (e.g., Autor et al. 2017).

  This section brings together those chapters that emphasize (largely macro-

  economic) ideas related to growth, inequality, and jobs. If the impact of

  AI will be like these other technologies, then what will the consequences

  8 Ajay Agrawal, Joshua Gans, and Avi Goldfarb

  look like for inequality, political economy, economic growth, jobs, and the

  meaning of work?

  Stevenson (chapter 7) outlines many of the key issues. She emphasizes that

  economists generally agree that in the long run society will be wealthier. She

  highlights issues with respect to the short run and income distribution. Sum-

  marizing both the tension in the public debate and the key themes in several

  other chapters, she notes, “In the end, there’s really two separate questions:

  there’s 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’s

  an income question, can we fi nd a stable and fair distribution of income?”

  Acemoglu and Restrepo (chapter 8) examine how AI and automation

  might change the nature of work. They suggest a task- based approach to un-

  derstanding automation, emphasizing the relative roles of labor and capital

  in the economy. “At the heart of our framework is the idea that automation

  and thus AI and robotics replace workers in tasks that they previously per-

  formed, and via this channel, create a powerful displacement eff ect.” This

  will lead to a lower labor share of economic output. At the same time, pro-

  ductivity will increase and capital will accumulate, thereby increasing the

  demand for labor. More importantly, “we argue that there is a more power-

  ful countervailing force that increases the demand for labor as well as the

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

  and activities in which labor has a comparative advantage relative to ma-

  chines. The creation of new tasks generates a reinstatement eff ect directly

  counterbalancing the displacement eff ect.” Like Stevenson, the long- run

  message is optimistic; however, a key point is that adjustment costs may be

  high. New skills are a necessary condition of the long- run optimistic fore-

  cast, and there is likely to be a short- and medium- term mismatch between

  skills and technologies. They conclude with a discussion of open questions

  about which skills are needed, the political economy of technological change

  (reinforcing ideas highlighted in the earlier chapter by Trajtenberg), and

  the interaction between inequality and the type of innovation enabled by

  automation going forward.

  Aghion, Jones, and Jones (chapter 9) build on the task- based model,

  focusing on the impact on economic growth. They emphasize Baumol’s

  cost disease: “Baumol (1967) observed that sectors with rapid productivity

  growth, such as agriculture and even manufacturing today, often see their

  share of GDP decline while those sectors with relatively slow productiv-

  ity growth—perhaps including many services—experience increases. As a

  consequence, economic growth may be constrained not by what we do well,

  but rather by what is essential and yet hard to improve. We suggest that com-

  bining this feature of growth with automation can yield a rich description of

  the growth process, including consequences for future growth and income

  distribution.” Thus, even in the limit where there is an artifi cial general

  intelligence that creates a singularity or intelligence explosion with a self-

  9

  improving AI, cost disease forces may constrain growth. This link between

  technological advance and Baumol’s cost disease provides a fundamental

  limit to the most optimistic and the most pessimistic views. Scarcity limits

  both growth and the downside risk. The chapter also explores how AI might

  reduce economic growth if it makes it easier to imitate a rival’s innovations,

  returning to issues of intellectual property highlighted in Mitchell’s com-

  ment. Finally, they discuss inequality within and across fi rms. They note

  that AI will increase wages of the least skilled employees of technologically

  advanced fi rms, but also increasingly outsource the tasks undertaken by

  such employees.

  Francois’s comment takes this emphasis on cost disease as a starting

  point, asking what those tasks will be that humans are left to do. “But it

  is when we turn to thinking about what are the products or services where

  humans will remain essential in production that we start to run into prob-

  lems. What if humans can’t do anything better than machines? Many dis-

  cussions at the conference centered around this very possibility. And I must

  admit that I found the scientists’ views compelling on this. . . . The point

  I wish to make is that even in such a world where machines are better at

  all tasks, there will still be an important role for human ‘work.’ And that

  work will become the almost political task of managing the machines.” He

  argues that humans must tell the machines what to optimize. Bostrom (2014)

  describes this as the value- loading problem. Francois emphasizes that this

  is largely a political problem, and links the challenges in identifying values

  with Arrow’s ([1951] 1963) impossibility theorem. He identifi es key ques-

  tions around ownership of the machines, length of time that rents should

  accrue to those owners, and the political structure of decision- making. In

  raising these questions, he provides a diff erent perspective on issues high-

  lighted by Stevenson on the meaning of work and Trajtenberg on the po-

  litical economy of technological change.

  The discussion of the meaning of work is a direct consequence of con-

  cerns about the impact of AI on jobs. Jobs have been the key focus of public

  discussion on AI and the economy. If human tasks get automated, what is

  left for humans to do? Bessen (chapter 10) explores this question, using data />
  about other technological advances to support his arguments. He empha-

  sizes that technological change can lead to an increase in demand and so

  the impact of automation on jobs is ambiguous, even within a sector. “The

  reason automation in textiles, steel, and automotive manufacturing led to

  strong job growth has to do with the eff ect of technology on demand. . . .

  New technologies do not just replace labor with machines, but in a com-

  petitive market, automation will reduce prices. In addition, technology may

  improve product quality, customization, or speed of delivery. All of these

  things can increase demand. If demand increases suffi

  ciently, employment

  will grow even though the labor required per unit of output declines.”

  Like Bessen, Goolsbee (chapter 11) notes that much of the popular dis-

  10 Ajay Agrawal, Joshua Gans, and Avi Goldfarb

  cussion around AI relates to labor market consequences. Recognizing that

  those consequences matter, his chapter mostly emphasizes the positive:

  growth and productivity are good. Artifi cial Intelligence has potential to

  increase our standard of living. Like Acemoglu and Restrepo, he notes that

  the short- term displacement eff ects could be substantial. One frequently

  cited solution to the displacement eff ects of AI is a universal basic income,

  in which all members of society receive a cash transfer from the government.

  He then discusses the economics of such a policy and the numerous chal-

  lenges to making it work. “First . . . in a world where AI- induced unemploy-

  ment is already high, separating work and income is an advantage. In a world

  like the one we are in now, off ering a basic income will likely cause a sizable

  drop in the labor market participation by low- wage groups. . . . Second,

  for a given amount of money to be used on redistribution, UBI likely shifts

  money away from the very poor. . . . Third, . . . converting things to a UBI

  and getting rid of the in-kind safety net will lead to a situation in which,

  even if among a small share of UBI recipients, SOME people will blow their

  money in unsympathetic ways—gambling, drugs, junk food, Ponzi schemes,

  whatever. And now those people will come to the emergency room or their

  kids will be hungry and by the rules, they will be out of luck. That’s what they

  were supposed to have used their UBI for.” Before concluding, he touches

  on a variety of regulatory issues that receive more detailed discussion in

  chapters 16 through 20. His conclusion mirrors that of Francois, emphasiz-

  ing the importance of humans in determining policy direction, even if AI

  improves to the point where it surpasses human intelligence.

  Furman (chapter 12) is similarly optimistic, emphasizing that we need

  more, not less AI. “AI is a critical area of innovation in the U.S. economy

  right now. At least to date, AI has not had a large impact on the aggregate

  performance of the macroeconomy or the labor market. But it will likely

  become more important in the years to come, bringing substantial oppor-

  tunities – and our fi rst impulse should be to embrace it fully.” Referencing

  data on productivity growth and on the diff usion of industrial robots, he

  then discusses potential negative eff ects on the economy as AI diff uses, par-

  ticularly with respect to inequality and reduced labor force participation.

  The issues around labor force participation highlight the importance of Ste-

  venson’s questions on the meaning of work. Like Goolsbee, Furman notes

  several challenges to implementing a universal basic income as a solution

  to these negative eff ects. He concludes that policy has an important role to

  play in enabling society to fully reap the benefi ts of technological change

  while minimizing the disruptive eff ects.

  Returning to the question of labor share highlighted by Acemoglu and

  Restrepo, Sachs (chapter 13) emphasizes that the income share going to

  capital grows with automation: “Rather than Solow- era stylized facts, I

  would therefore propose the following alternative stylized facts: (a) the

  share of national income accruing to capital rises over time in sectors expe-

  11

  riencing automation, especially when capital is measured to include human

  capital; (b) the share of national income accruing to low- skill labor drops

  while the share accruing to high- skill labor rises; (c) the dynamics across sec-

  tors vary according to the diff erential timing of automation, with automa-

  tion spreading from low- skilled and predictable tasks toward high- skilled

  and less predictable tasks; (d) automation refl ects the rising intensity of

  science and technology throughout the economy . . ., and (e) future techno-

  logical changes associated with AI are likely to shift national income from

  medium- skilled and high- skilled toward owners of business capital.” The

  chapter concludes with a list of key open questions about the dynamics of

  auto mation, the role of monopoly rents, and the consequences for income

  distribution and labor force participation.

  Korinek and Stiglitz (chapter 14) also emphasize income distribution,

  discussing the implications of AI- related innovation for inequality. They

  show that, in a fi rst- best economy, contracts can be specifi ed in advance that

  make innovation Pareto improving. However, imperfect markets and costly

  redistribution can imply a move away from the fi rst- best. Innovation may

  then drive inequality directly by giving innovators a surplus, or indirectly

  by changing the demand for diff erent types of labor and capital. They dis-

  cuss policies that could help reduce the increase in inequality, emphasizing

  diff erent taxation tools. Related to the ideas introduced in Mitchell’s com-

  ment, they also explore IP policies: “If outright redistribution is infeasible,

  there may be other institutional changes that result in market distributions

  that are more favorable to workers. For example, intervention to steer tech-

  nological progress may act as a second- best device . . . we provide an ex-

  ample in which a change in intellectual property rights—a shortening of the

  term of patent protection—eff ectively redistributes some of the innovators’

  surplus to workers (consumers) to mitigate the pecuniary externalities on

  wages that they experience, with the ultimate goal that the benefi ts of the

  innovation are more widely shared.” Stiglitz and Korinek conclude with a

  more speculative discussion of artifi cial general intelligence (superhuman

  artifi cial intelligence), emphasizing that such a technological development

  will likely further increase inequality.

  The fi nal chapter in the section on growth, jobs, and inequality calls for

  a diff erent emphasis. Cowen (chapter 15) emphasizes consumer surplus,

  international eff ects, and political economy. With respect to consumer sur-

  plus, he writes, “Imagine education and manufactured goods being much

  cheaper because we produced them using a greater dose of smart software.

  The upshot is that even if a robot puts you out of a job or lowers your pay,

  there will be some recompense on the consumer side.” Cowen also specu-

  lates that AI
might hurt developing countries much more than developed,

  as automation means that labor cost reasons to off shore decline. Finally,

  like Trajtenberg and Francois, he emphasizes the political economy of AI,

  highlighting questions related to income distribution.

  12 Ajay Agrawal, Joshua Gans, and Avi Goldfarb

  Taken together, the chapters in this section highlight several key issues and

  provide models that identify challenges related to growth, jobs, inequality,

  and politics. These models set up a number of theoretical and empirical

  questions about how AI will impact economic outcomes within and across

  countries.

  The discussions are necessarily speculative because AI has not yet diff used

  widely, so research must either be entirely theoretical or it must use related

  technologies (such as factory robots) as a proxy for AI. The discussions are

  also speculative because of the challenges in measuring the relevant vari-

  ables. In order to determine the impact of AI on the economy, we need con-

  sistent measures of AI, productivity, intangible capital, and growth across

  sectors, regions, and contexts. Going forward, to the extent that progress

  occurs against the proposed research agenda, it will depend on advances

  in measurement.

  Machine Learning and Regulation

  Industry will be a key innovator and adopter of artifi cial intelligence.

  A number of regulatory issues arise. The regulatory issues related to truly

  intelligent machines are touched on by Trajtenberg, Francois, Goolsbee, and

  Cowen. Mitchell’s comment of Cockburn, Henderson, and Stern empha-

  sizes intellectual property regulation. This section focuses on other regula-

  tory challenges with respect to advances in machine learning.

  Varian (chapter 16) sets up the issues by describing the key models from

  industrial organization that are relevant to understanding the impact of

  machine learning on fi rms. He highlights the importance of data as a scarce

  resource, and discusses the economics of data as an input: it is nonrival and

  it exhibits decreasing returns to scale in a technical sense (because predic-

  tion accuracy increases in the square root of N ). He discusses the structure

  of ML- using industries including vertical integration, economies of scale,