The Economics of Artificial Intelligence Read online

  discussions, and debates, participants weighed in with their ideas for what

  the key questions will be, what research has already shown, and where the

  challenges will lie. Pioneering AI researchers Geoff rey Hinton, Yann LeCun,

  and Russ Salakhutdinov attended, providing useful context and detail about

  the current and expected future capabilities of the technology. The confer-

  ence was unique because it emphasized the work that still needs to be done,

  rather than the presentation of standard research papers. Participants had

  the freedom to engage in informed speculation and healthy debate about the

  most important areas of inquiry.

  This volume contains a summary of the proceedings of the conference.

  We provided authors with few constraints. This meant diversity in topics and

  chapter style. Many of the chapters contained herein are updated versions

  of the original papers and presentations at the conference. Some discussants

  commented directly on the chapters while others went further afi eld, empha-

  sizing concepts that did not make it into the formal presentations but instead

  arose as part of debate and discussion. The volume also contains a small

  number of chapters that were not presented at the conference, but never-

  theless represent ideas that came up in the general discussion and that war-

  ranted inclusion in a volume describing the proceedings of the conference.

  We categorize the chapters into four broad themes. First, several chapters

  emphasize the role of AI as a general purpose technology (GPT), building

  on the existing literature on general purpose technologies from the steam

  engine to the internet. Second, many chapters highlight the impact of AI

  on growth, jobs, and inequality, focusing on research and tools from macro

  and labor economics. Third, fi ve chapters discuss machine learning and eco-

  nomic regulation, with an emphasis on microeconomic consequences and

  industrial organization. The fi nal set of chapters explores how AI will aff ect

  research in economics.

  Of course, these themes are not mutually exclusive. Discussion of AI as

  a GPT naturally leads to discussions of economic growth. Regulation can

  enhance or reduce inequality. And AI’s impact on economics is a conse-

  quence of it being a general purpose technology for scientifi c discovery (as

  emphasized in chapter 4 by Cockburn, Henderson, and Stern). Further-

  more, a handful of concepts cut across the various parts, most notably the

  3

  role of humans as AI improves and the interaction between technological

  advance and political economy.

  Below, we summarize these four broad themes in detail. Before doing so,

  we provide a defi nition of the technology that brings together the various

  themes.

  What Is Artifi cial Intelligence?

  The Oxford English Dictionary defi nes artifi cial intelligence as “the

  theory and development of computer systems able to perform tasks nor-

  mally requiring human intelligence.” This defi nition is both broad and fl uid.

  There is an old joke among computer scientists that artifi cial intelligence

  defi nes what machines cannot yet do. Before a machine could beat a human

  expert at chess, such a win would mean artifi cial intelligence. After the famed

  match between IBM’s Deep Blue and Gary Kasparov, playing chess was

  called computer science and other challenges became artifi cial intelligence.

  The chapters in this volume discuss three related, but distinct, concepts

  of artifi cial intelligence. First, there is the technology that has driven the

  recent excitement around artifi cial intelligence: machine learning. Machine

  learning is a branch of computational statistics. It is a tool of prediction in

  the statistical sense, taking information you have and using it to fi ll in infor-

  mation you do not have. Since 2012, the uses of machine learning as a pre-

  diction technology have grown substantially. One set of machine- learning

  algorithms, in particular, called “deep learning,” has been shown to be useful

  and commercially viable for a variety of prediction tasks from search engine

  design to image recognition to language translation. The chapter in the book

  authored by us—Agrawal, Gans, and Goldfarb—emphasizes that rapid

  improvements in prediction technology can have a profound impact on orga-

  nizations and policy (chapter 3). The chapter by Taddy (chapter 2) defi nes

  prediction with machine learning as one component of a true artifi cial intel-

  ligence and provides detail on the various machine- learning technologies.

  While the recent interest in AI is driven by machine learning, computer

  scientists and philosophers have emphasized the feasibility of a true artifi -

  cial general intelligence that equals or exceeds human intelligence (Bostrom

  2014; Kaplan 2016). The closing sentence of this volume summarizes this

  possibility bluntly. Daniel Kahneman writes, “I do not think that there is

  very much that we can do that computers will not eventually be programmed

  to do.” The economic and societal impact of machines that surpass human

  intelligence would be extraordinary. Therefore—whether such an event

  occurs imminently, in a few decades, in a millennium, or never—it is worth

  exploring the economic consequences of such an event. While not a focal

  aspect of any chapter, several of the chapters in this volume touch on the

  economic consequences of such superintelligent machines.

  A third type of technology that is often labeled “artifi cial intelligence” is

  4 Ajay Agrawal, Joshua Gans, and Avi Goldfarb

  better seen as a process: automation. Much of the existing empirical work on

  the impact of artifi cial intelligence uses data on factory automation through

  robotics. Daron Acemoglu and Pascual Restrepo use data on factory robots

  to explore the impact of AI and automation on work (chapter 8). Auto-

  mation is a potential consequence of artifi cial intelligence, rather than arti-

  fi cial intelligence per se. Nevertheless, discussions of the consequences of

  artifi cial intelligence and automation are tightly connected.

  While most chapters in the book focus on the fi rst defi nition—artifi cial

  intelligence as machine learning—a prediction technology, the economic

  implications of artifi cial general intelligence and automation receive seri-

  ous attention.

  AI as a GPT

  A GPT is characterized by pervasive use in a wide range of sectors com-

  bined with technological dynamism (Bresnahan and Trajtenberg 1995).

  General purpose technologies are enabling technologies that open up new

  opportunities. While electric motors did reduce energy costs, the productiv-

  ity impact was largely driven by increased fl exibility in the design and loca-

  tion of factories (David 1990). Much of the interest in artifi cial intelligence

  and its impact on the economy stems from its potential as a GPT. Human

  intelligence is a general purpose tool. Artifi cial intelligence, whether defi ned

  as prediction technology, general intelligence, or automation, similarly has

  potential to apply across a broad range of sectors.

  Brynjolfsson, Rock, and Syverson (chapter 1) argue the case for
AI as a

  GPT. They focus on machine learning and identify a variety of sectors in

  which machine learning is likely to have a broad impact. They note expected

  continual technological progress in machine learning and a number of com-

  plementary innovations that have appeared along with machine learning.

  By establishing AI as a GPT, they can turn to the general lessons of the pro-

  ductivity literature on GPTs with respect to initially low rates of productiv-

  ity growth, organizational challenges, and adjustment costs. They propose

  four potential explanations for the surprisingly low measured productivity

  growth given rapid innovation in AI and related technologies—false hopes,

  mismeasurement, redistribution, and implementation lags—and conclude

  that lags due to missing complementary innovations are most likely the

  primary source of missing productivity growth: “an underrated area of

  research involves the complements to the new AI technologies, not only

  in areas of human capital and skills, but also new processes and business

  models. The intangible assets associated with the last wave of computeriza-

  tion were about ten times as large as the direct investments in computer

  hardware itself.”

  Henderson’s comment emphasizes the impact of a GPT on employment

  and the distribution of income, directly linking the discussion of AI as a

  5

  GPT to questions addressed in the section on Growth, Jobs, and Inequal-

  ity. She agrees with the central thesis “One of the reasons I like the paper

  so much is that it takes seriously an idea that economists long resisted—

  namely, that things as nebulous as ‘culture’ and ‘organizational capabilities’

  might be (a) very important, (b) expensive, and (c) hard to change.” At the

  same time, she adds emphasis on additional implications: “I think that the

  authors may be underestimating the implications of this dynamic in impor-

  tant ways. . . . I’m worried about the transition problem at the societal level

  quite as much as I’m worried about it at the organizational level.”

  The next chapters provide micro- level detail on the nature of AI as a

  technology. Taddy (chapter 2) provides a broad overview of the meaning

  of intelligence in computer science. He then provides some technical detail

  on two key machine- learning techniques, deep learning and reinforcement

  learning. He explains the technology in a manner intuitive to economists:

  “Machine learning is a fi eld that thinks about how to automatically build

  robust predictions from complex data. It is closely related to modern statis-

  tics, and indeed many of the best ideas in ML have come from statisticians

  (the lasso, trees, forests, etc.). But whereas statisticians have often focused

  on model inference—on understanding the parameters of their models (e.g.,

  testing on individual coeffi

  cients in a regression)—the ML community has

  been more focused on the single goal of maximizing predictive performance.

  The entire fi eld of ML is calibrated against ‘out- of-sample’ experiments that

  evaluate how well a model trained on one data set will predict new data.”

  Building on ideas in Agrawal, Gans, and Goldfarb (2018), we argue in

  chapter 3 that the current excitement around AI is driven by advances in

  prediction technology. We then show that modeling AI as a drop in the cost

  of prediction provides useful insight into the microeconomic impact of AI

  on organizations. We emphasize that AI is likely to substitute for human

  prediction, but complement other skills such as human judgment—defi ned

  as knowing the utility or valuation function: “a key departure from the

  usual assumptions of rational decision- making is that the decision- maker

  does not know the payoff from the risky action in each state and must apply

  judgment to determine the payoff . . . . Judgment does not come for free.”

  Prat’s comment emphasizes that economists typically assume that the

  valuation function is given, and that loosening that assumption will lead to

  a deeper understanding of the impact of AI on organizations. He off ers an

  example to illustrate: “Admissions offi

  ces of many universities are turning to

  AI to choose which applicants to make off ers to. Algorithms can be trained

  on past admissions data. We observe the characteristics of applicants and

  the grades of past and present students. . . . The obvious problem is that we

  do not know how admitting someone who is likely to get high grades is going

  to aff ect the long- term payoff of our university. . . . Progress in AI should

  induce our university leaders to ask deeper questions about the relationship

  between student quality and the long- term goals of our higher- learning

  6 Ajay Agrawal, Joshua Gans, and Avi Goldfarb

  institutions. These questions cannot be answered with AI, but rather with

  more theory- driven retrospective approaches or perhaps more qualitative

  methodologies.”

  The next chapters explore AI as a GPT that will enhance science and

  innovation. After reviewing the history of artifi cial intelligence, Cockburn,

  Henderson, and Stern (chapter 4) provide empirical support for the wide-

  spread application of machine learning in general, and deep learning in

  particular, in scientifi c fi elds outside of computer science: “we develop what

  we believe is the fi rst systematic database that captures the corpus of scien-

  tifi c paper and patenting activity in artifi cial intelligence . . . we fi nd striking evidence for a rapid and meaningful shift in the application orientation of

  learning- oriented publications, particularly after 2009.” The authors make

  a compelling case for AI as a general purpose tool in the method of inven-

  tion. The chapter concludes by discussing the implications for innovation

  policy and innovation management: “the potential commercial reward from

  mastering this mode of research is likely to usher in a period of racing,

  driven by powerful incentives for individual companies to acquire and con-

  trol critical large data sets and application- specifi c algorithms.”

  Mitchell’s comment emphasizes the regulatory eff ects of AI as a GPT

  for science and innovation—in terms of intellectual property, privacy, and

  competition policy: “It is not obvious whether AI is a general purpose tech-

  nology for innovation or a very effi

  cient method of imitation. The answer

  has a direct relevance for policy. A technology that made innovation cheaper

  would often (but not always) imply less need for strong IP protection, since

  the balance would swing toward limiting monopoly power and away from

  compensating innovation costs. To the extent that a technology reduces

  the cost of imitation, however, it typically necessitates greater protection.”

  Several later chapters detail these and other regulatory issues.

  Agrawal, McHale, and Oettl (chapter 5) provide a recombinant growth

  model that explores how a general purpose technology for innovation could

  aff ect the rate of scientifi c discovery: “instead of emphasising the potential

  substitution of machines for workers in existing tasks, we emphasise the

  importance of A
I in overcoming a specifi c problem that impedes human

  researchers—fi nding useful combinations in complex discovery spaces . . .

  we develop a relatively simple combinatorial- based knowledge production

  function that converges in the limit to the Romer/ Jones function. . . . If the

  curse of dimensionality is both the blessing and curse of discovery, then

  advances in AI off er renewed hope of breaking the curse while helping to

  deliver on the blessing.” This idea of AI as an input into innovation is a

  key component of Cockburn, Henderson, and Stern (chapter 4), as well as

  in several later chapters. It is an important element of Aghion, Jones, and

  Jones’s model of the impact of AI on economic growth (chapter 9), empha-

  sizing endogenous growth through AI (self-)improvements. It also underlies

  7

  the chapters focused on how AI will impact the way economics research is

  conducted (chapters 21 through 24).

  The section on AI as a general purpose technology concludes with Manuel

  Trajtenberg’s discussion of political and societal consequences (chapter 6).

  At the conference, Trajtenberg discussed Joel Mokyr’s paper “The Past and

  Future of Innovation: Some Lessons from Economic History,” which will

  be published elsewhere. The chapter therefore sits between a stand- alone

  chapter and a discussion. Trajtenberg’s chapter does not comment directly

  on Mokyr, but uses Mokyr’s paper as a jumping- off point to discuss how

  technology creates winners and losers, and the policy challenges associated

  with the political consequences of the diff usion of a GPT. “The sharp split

  between winners and losers, if left to its own, may have serious consequences

  far beyond the costs for the individuals involved: when it coincides with the

  political divide, it may threaten the very fabric of democracy, as we have seen

  recently both in America and in Europe. Thus, if AI bursts onto the scene

  and triggers mass displacement of workers, and demography plays out its

  fateful hand, the economy will be faced with a formidable dual challenge,

  that may require a serious reassessment of policy options . . . we need to

  anticipate the required institutional changes, to experiment in the design

  of new policies, particularly in education and skills development, in the

  professionalization of service occupations, and in aff ecting the direction of