The Economics of Artificial Intelligence Read online

  and the potential for price discrimination. He emphasizes the diff erence

  between learning by doing and data network eff ects: “There is a concept that

  is circulating among lawyers and regulators called ‘data network eff ects.’ The

  model is that a fi rm with more customers can collect more data and use this

  data to improve its product. This is often true—the prospect of improving

  operations is what makes ML attractive—but it is hardly novel. And it is

  certainly not a network eff ect! This is essentially a supply- side eff ect known

  as ‘learning by doing.’. . . A company can have huge amounts of data, but

  if it does nothing with the data, it produces no value. In my experience, the

  problem is not lack of resources, but is lack of skills. A company that has

  data but no one to analyze it is in a poor position to take advantage of that

  data.” He concludes by highlighting policy questions related to algorithmic

  collusion (which was discussed at the conference as “economist catnip,”

  13

  interesting and fun but unlikely to be of fi rst- order importance), security,

  privacy, and transparency.

  Chevalier’s comment builds on Varian’s emphasis on the importance of

  data, exploring the potential of antitrust policy aimed at companies that

  use machine learning. Legal scholars and policymakers have asked whether

  antitrust essential facilities doctrine should be applied to data ownership.

  She emphasizes the trade- off between static and dynamic considerations for

  such a policy: “In evaluating antitrust policies in innovative industries, it is

  important to recognize that consumer benefi ts from new technologies arise

  not just from obtaining goods and services at competitive prices, but also

  from the fl ow of new and improved products and services that arise from

  innovation. Thus, antitrust policy should be evaluated not just in terms of

  its eff ect on prices and outputs, but also on its eff ect on the speed of inno-

  vation. Indeed, in the high technology industries, it seems likely that these

  dynamic effi

  ciency considerations dwarf the static effi

  ciency considerations.”

  She also explores several practical challenges.

  Another regulatory issue that arises from the importance of data is pri-

  vacy. Tucker (chapter 17) notes that machine learning uses data to make

  predictions about what individuals may desire, be infl uenced by, or do. She

  emphasizes that privacy is challenging for three reasons: cheap storage

  means that data may persist longer than the person who generated the data

  intended, nonrivalry means that data may be repurposed for uses other than

  originally intended, and externalities caused by data created by one indi-

  vidual that contains information about others: “For example, in the case of

  genetics, the decision to create genetic data has immediate consequences for

  family members, since one individual’s genetic data is signifi cantly similar to

  the genetic data of their family members. . . . There may also be spillovers

  across a person’s decision to keep some information secret, if such secrecy

  predicts other aspects of that individual’s behavior that AI might be able

  to project from.” She discusses potential negative impacts of these three

  challenges, concluding with some key open questions.

  Jin (chapter 18) also focuses on the importance of data as an input into

  machine learning. She emphasizes that reduced privacy creates security

  challenges, such as identity theft, ransomware, and misleading algorithms

  (such as Russian- sponsored posts in the 2016 US election): “In my opinion,

  the leading concern is that fi rms are not fully accountable for the risk they

  bring to consumer privacy and data security. To restore full accountability,

  one needs to overcome three obstacles, namely (a) the diffi

  culty to observe

  fi rms’ actual action in data collection, data storage, and data use; (b) the

  diffi

  culty to quantify the consequence of data practice, especially before low-

  probability adverse events realize themselves; and (c) the diffi

  culty to draw a

  causal link between a fi rm’s data practice and its consequence.” Combined,

  Tucker and Jin’s chapters emphasize that any discussion of growth and

  14 Ajay Agrawal, Joshua Gans, and Avi Goldfarb

  impact of AI requires an understanding of the privacy framework. Access

  to data drives innovation, underlies the potential for economic growth, and

  frames the antitrust debate.

  The economics of data also create challenges with respect to the rules

  governing international trade. Goldfarb and Trefl er (chapter 19) argue that

  economies of scale in data through feedback loops, along with economies

  of scope and knowledge externalities in AI innovation, could create the

  opportunity for country- level rents and strategic trade policy. At the same

  time, they emphasize that the geographic constraints on data and knowledge

  would have to be high for such a policy to be optimal at the country level.

  They highlight the rise of China: “China has become the focal point for

  much of the international discussion. The US narrative has it that Chinese

  protection has reduced the ability of dynamic US fi rms such as Google and

  Amazon to penetrate Chinese markets. This protection has allowed China

  to develop signifi cant commercial AI capabilities, as evidenced by compa-

  nies such as Baidu (a search engine like Google), Alibaba (an e-commerce

  web portal like Amazon), and Tencent (the developer of WeChat, which

  can be seen as combining the functions of Skype, Facebook, and Apple

  Pay) . . . we collected time- series data on the institutional affi

  liation of all

  authors of papers presented at a major AI research conference . . . we com-

  pare the 2012 and 2017 conferences. . . . While these countries all increased

  their absolute number of participants, in relative terms they all lost ground

  to China, which leapt from 10 percent in 2012 to 23 percent in 2017.” The

  authors discuss the international dimensions of domestic regulation related

  to privacy, access to government data, and industrial standards.

  The fi nal regulatory issue highlighted in this section is tort liability.

  Galasso and Luo (chapter 20) review prior literature on the relationship

  between liability and innovation. They emphasize the importance of getting

  the balance right between consumer protection and innovation incentives:

  “A central question in designing a liability system for AI technologies is

  how liability risk should be allocated between producers and consumers,

  and how this allocation might aff ect innovation. . . . A key promise of AI

  technologies is to achieve autonomy. With less room for consumers to take

  precautions, the relative liability burden is likely to shift toward producers,

  especially in situations in which producers are in a better position than indi-

  vidual users to control risk. . . . On the other hand, during the transitional

  period of an AI technology, substantial human supervision may still be

  required. . . . In many of these situations, it may be impractical or too costly


  for producers to monitor individual users and to intervene. Therefore, it

  would be important to maintain consumer liability to the extent that users

  of AI technologies have suffi

  cient incentives to take precautions and invest

  in training, thus internalizing potential harm to others.”

  Broadly, regulation will aff ect the speed at which AI diff uses. Too much

  regulation, and industry will not have incentives to invest. Too little regu-

  15

  lation, and consumers will not trust the products that result. In this way,

  getting the regulatory balance right is key to understanding when and how

  any impact of AI on economic growth and inequality will arise.

  Impact on the Practice of Economics

  Cockburn, Henderson, and Stern emphasize that machine learning is a

  general purpose technology for science and innovation. As such, it is likely

  to have an impact on research in a variety of disciplines, including eco-

  nomics. Athey (chapter 21) provides an overview of the various ways in

  which machine learning is likely to aff ect the practice of economics. For

  example: “I believe that machine learning (ML) will have a dramatic impact

  on the fi eld of economics within a short time frame. . . . ML does not add

  much to questions about identifi cation, which concern when the object of

  interest, for example, a causal eff ect, can be estimated with infi nite data, but

  rather yields great improvements when the goal is semiparametric estima-

  tion or when there are a large number of covariates relative to the number

  of observations . . . a key advantage of ML is that ML views empirical

  analysis as “algorithms” that estimate and compare many alternative mod-

  els . . . ‘outsourcing’ model selection to algorithms works very well when

  the problem is ‘simple’—for example, prediction and classifi cation tasks,

  where performance of a model can be evaluated by looking at goodness of

  fi t in a held- out test set.” She emphasizes the usefulness of machine- learning

  techniques for policy problems related to prediction (as in Kleinberg et al.

  2015). The chapter then details recent advances in using machine- learning

  techniques in causal inference, which she views as a fundamental new tool kit

  for empirical economists. She concludes with a list of sixteen predictions of

  how machine learning will impact economics, emphasizing new econometric

  tools, new data sets and measurement techniques, increased engagement of

  economists as engineers (and plumbers), and, of course, increased study

  of the economic impact of machine learning on the economy as a whole.

  Lederman’s comment emphasizes the usefulness of machine learning to

  create new variables for economic analysis, and how the use of machine

  learning by organizations creates a new kind of endogeneity problem: “We

  develop theoretical models to help us understand the data- generation pro-

  cess which, in turn, informs both our concerns about causality as well as

  the identifi cation strategies we develop. . . . Overall, as applied researchers

  working with real- world data sets, we need to recognize that increasingly

  the data we are analyzing is going to be the result of decisions that are made

  by algorithms in which the decision- making process may or may not re-

  semble the decision- making processes we model as social scientists.”

  If the study of AI is going to be a key question for economists going for-

  ward, Raj and Seamans (chapter 22) emphasize that we need better data:

  “While there is generally a paucity of data examining the adoption, use, and

  16 Ajay Agrawal, Joshua Gans, and Avi Goldfarb

  eff ects of both AI and robotics, there is currently less information available

  regarding AI. There are no public data sets on the utilization or adoption of

  AI at either the macro or micro level. The most complete source of informa-

  tion, the McKinsey Global Institute study, is proprietary and inaccessible

  to the general public or the academic community. The most comprehensive

  and widely used data set examining the diff usion of robotics is the Inter-

  national Federation of Robotics (IFR) Robot Shipment Data . . . the IFR

  does not collect any information on dedicated industrial robots that serve

  one purpose. Furthermore, some of the robots are not classifi ed by indus-

  try, detailed data is only available for industrial robots (and not robots in

  service, transportation, warehousing, or other sectors), and geographical

  information is often aggregated” They provide a detailed discussion of data-

  collection opportunities by government and by academic researchers. If the

  agenda set up in the other chapters is to be answered, it is important to have

  a reliable data set that defi nes AI, measures its quality, and tracks its diff usion.

  Related to Athey’s emphasis of increased engagement of economists

  as engineering, Milgrom and Tadelis (chapter 23) describe how machine

  learning is already aff ecting market- design decisions. Using specifi c ex-

  amples from online marketplaces and telecommunications auctions, they

  emphasize the potential of AI to improve effi

  ciency by predicting demand

  and supply, overcoming computational barriers, and reducing search fric-

  tions: “AI and machine learning are emerging as important tools for market

  design. Retailers and marketplaces such as eBay, Taobao, Amazon, Uber,

  and many others are mining their vast amounts of data to identify patterns

  that help them create better experiences for their customers and increase

  the effi

  ciency of their markets . . . two- sided markets such as Google, which

  match advertisers with consumers, are not only using AI to set reserve prices

  and segment consumers into fi ner categories for ad targeting, but they also

  develop AI- based tools to help advertisers bid on ads. . . . Another impor-

  tant application of AI’s strength in improving forecasting to help markets

  operate more effi

  ciently is in electricity markets. To operate effi

  ciently, elec-

  tricity market makers . . . must engage in demand and supply forecasting.”

  The authors argue that AI will play a substantial role in the design and

  implementation of markets over a wide range of applications.

  Camerer (chapter 24) also emphasizes the role of AI as a tool for predict-

  ing choice: “Behavioral economics can be defi ned as the study of natural

  limits on computation, willpower, and self- interest, and the implications of

  those limits for economic analysis (market equilibrium, IO, public fi nance,

  etc.). A diff erent approach is to defi ne behavioral economics more generally,

  as simply being open- minded about what variables are likely to infl uence

  economic choices. . . . In a general ML approach, predictive features could

  be—and should be—any variables that predict. . . . If behavioral econom-

  ics is recast as open- mindedness about what variables might predict, then

  ML is an ideal way to do behavioral economics because it can make use of

  17

  a wide set of variables and select which ones predict.” He argues that fi rms,

  policymakers, and market designers can implement AI as either a “bionic<
br />
  patch” that improves human decision- making or “malware” that exploits

  human weaknesses. In this way, AI could reduce or exacerbate the political

  economy and inequality issues highlighted in earlier chapters. In addition,

  Camerer explores two other ways in which AI and behavioral economics will

  interact. He hypothesizes that machine learning could help predict human

  behavior in a variety of settings including bargaining, risky choice, and

  games, helping to verify or reject theory. He also emphasizes that (poor)

  implementation of AI might provide insight into new ways to model biases

  in human decision- making.

  The book concludes with Kahneman’s brief and insightful comment.

  Kahneman begins with a discussion of Camerer’s idea of using prediction

  to verify theory, but continues with a broader discussion of a variety of

  themes that arose over the course of the conference. With an optimistic

  tone, he emphasizes that there are no obvious limits to what artifi cial intel-

  ligence may be able to do: “Wisdom is breadth. Wisdom is not having too

  narrow a view. That is the essence of wisdom; it is broad framing. A robot

  will be endowed with broad framing. When it has learned enough, it will

  be wiser than we people because we do not have broad framing. We are nar-

  row thinkers, we are noisy thinkers, and it is very easy to improve upon us.

  I do not think that there is very much that we can do that computers will

  not eventually be programmed to do.”

  The Future of Research on the Economics of Artifi cial Intelligence

  The chapters in this book are the beginning. They highlight key questions,

  recognize the usefulness of several economic models, and identify areas for

  further development. We can leverage what we know about GPTs to antici-

  pate the impact of AI as it diff uses, recognizing that no two GPTs are iden-

  tical. If AI is a general purpose technology, it is likely to lead to increased

  economic growth. A common theme in these chapters is that slowing down

  scientifi c progress—even if it were possible—would come at a signifi cant

  cost. At the same time, many attendees emphasized that the distribution

  of the benefi ts of AI might not be even. It depends on who owns the AI,

  the eff ect on jobs, and the speed of diff usion.

  The task given to the conference presenters was to scope out the research