The Economics of Artificial Intelligence

Home > Other > The Economics of Artificial Intelligence > Page 20
The Economics of Artificial Intelligence Page 20

by Ajay Agrawal


  erally complements, as long as judgment is not too diffi

  cult. We also show

  that improvements in judgment change the type of prediction quality that

  is most useful: better judgment means that more accurate predictions are

  valuable relative to more frequent predictions. Finally, we explore the role of

  complexity, demonstrating that, in the presence of complexity, the impact of

  improved prediction on the value of judgment depends on whether improved

  prediction leads to automated decision- making. Complexity is a key aspect

  of economic research in automation, contracting, and the boundaries of

  the fi rm. As prediction machines improve, our model suggests that the con-

  sequences in complex environments are particularly fruitful to study.

  There are numerous directions research in this area could proceed. First,

  the chapter does not explicitly model the form of the prediction—includ-

  ing what measures might be the basis for decision- making. In reality, this

  is an important design variable and impacts on the accuracy of predic-

  tions and decision- making. In computer science, this is referred to as the

  choice of surrogates, and this appears to be a topic amenable for economic

  theoretical investigation. Second, the chapter treats judgment as largely a

  human- directed activity. However, we have noted that it can else be encoded,

  but have not been explicit about the process by which this occurs. Endogenis-

  ing this—perhaps relating it to the accumulation of experience—would be

  an avenue for further investigation. Finally, this is a single- agent model. It

  would be interesting to explore how judgment and prediction mix when each

  is impacted upon by the actions and decisions of other agents in a game

  theoretic setting.

  References

  Acemoglu, Daron. 2003. “Labor- and Capital- Augmenting Technical Change.”

  Journal of the European Economic Association 1 (1): 1– 37.

  Acemoglu, Daron, and Pascual Restrepo. 2017. “The Race between Machine and

  Man: Implications of Technology for Growth, Factor Shares, and Employment.”

  Working paper, Massachusetts Institute of Technology.

  Agrawal, Ajay, Joshua S. Gans, and Avi Goldfarb. 2018a. “Human Judgment and

  AI Pricing.” American Economic Association: Papers & Proceedings, 108:58–63.

  ———. 2018b. Prediction Machines: The Simple Economics of Artifi cial Intelligence.

  Boston, MA: Harvard Business Review Press.

  Alpaydin, Ethem. 2010. Introduction to Machine Learning, 2nd ed. Cambridge, MA: MIT Press.

  Autor, David. 2015. “Why Are There Still So Many Jobs? The History and Future of

  Workplace Automation.” Journal of Economic Perspectives 29 (3): 3– 30.

  Baker, George, Robert Gibbons, and Kevin Murphy. 1999. “Informal Authority in

  Organizations.” Journal of Law, Economics, and Organization 15:56– 73.

  Belloni, Alexandre, Victor Chernozhukov, and Christian Hansen. 2014. “High-

  110 Andrea Prat

  Dimensional Methods and Inference on Structural and Treatment Eff ects.” Jour-

  nal of Economic Perspectives 28 (2): 29– 50.

  Benzell, Seth G., Laurence J. Kotlikoff , Guillermo LaGarda, and Jeff rey D. Sachs.

  2015. “Robots Are Us: Some Economics of Human Replacement.” NBER Work-

  ing Paper no. 20941, Cambridge, MA.

  Bolton, P., and A. Faure- Grimaud. 2009. “Thinking Ahead: The Decision Problem.”

  Review of Economic Studies 76:1205– 38.

  Brynjolfsson, Erik, and Andrew McAfee. 2014. The Second Machine Age. New York:

  W. W. Norton.

  Dogan, M., and P. Yildirim. 2017. “Man vs. Machine: When Is Automation Inferior

  to Human Labor?” Unpublished manuscript, The Wharton School of the Uni-

  versity of Pennsylvania.

  Domingos, Pedro. 2015. The Master Algorithm. New York: Basic Books.

  Forbes, Silke, and Mara Lederman. 2009. “Adaptation and Vertical Integration in

  the Airline Industry.” American Economic Review 99 (5): 1831– 49.

  Hastie, Trevor, Robert Tibshirani, and Jerome Friedman. 2009. The Elements of

  Statistical Learning: Data Mining, Inference, and Prediction, 2nd ed. New York: Springer.

  Hawkins, Jeff . 2004. On Intelligence. New York: Times Books.

  Jha, S., and E. J. Topol. 2016. “Adapting to Artifi cial Intelligence: Radiologists and Pathologists as Information Specialists.” Journal of the American Medical Association 316 (22): 2353– 54.

  Lusted, L. B. 1960. “Logical Analysis in Roentgen Diagnosis.” Radiology 74:178– 93.

  Markov, John. 2015. Machines of Loving Grace. New York: HarperCollins Pub-

  lishers.

  Ng, Andrew. 2016. “What Artifi cial Intelligence Can and Can’t Do Right Now.”

  Harvard Business Review Online. Accessed Dec. 8, 2016. https:// hbr .org/ 2016/ 11

  / what- artifi cial- intelligence- can- and- cant- do- right- now.

  Simon, H. A. 1951. “A Formal Theory of the Employment Relationship.” Econo-

  metrica 19 (3): 293– 305.

  Tadelis, S. 2002. “Complexity, Flexibility and the Make- or- Buy Decision.” American Economic Review 92 (2): 433– 37.

  Tirole, J. 2009. “Cognition and Incomplete Contracts.” American Economic Review

  99 (1): 265– 94.

  Varian, Hal R. 2014. “Big Data: New Tricks for Econometrics.” Journal of Economic

  Perspectives 28 (2): 3– 28.

  Comment Andrea Prat

  One of the key activities of organizations is to collect, process, combine,

  and utilize information (Arrow 1974). A modern corporation exploits

  the vast amounts of data that it accumulates from marketing, operations,

  human resources, fi nance, and other functions to grow faster and be more

  Andrea Prat is the Richard Paul Richman Professor of Business at Columbia Business

  School and professor of economics at Columbia University.

  For acknowledgments, sources of research support, and disclosure of the author’s material fi nancial relationships, if any, please see http:// www .nber .org/ chapters/ c14022.ack.

  Comment 111

  productive. This exploitation process depends on the kind of information

  technology (IT) that is available to the fi rm. If IT undergoes a revolution,

  we should expect deep structural changes in the way fi rms are organized

  (Milgrom and Roberts 1990).

  Agrawal, Gans, and Goldfarb explore the eff ects that an IT revolution

  centered on artifi cial intelligence could have on organizations. Their anal-

  ysis highlights an insightful distinction between prediction, the process of

  forecasting a state of the world given observable information, and judg-

  ment, the assessment of the eff ects of the state of the world and the possible action x the organization can take in response to it, namely, the value of the payoff function u(, x).

  This is an important point of departure from existing work. Almost all

  economists—as well as computer scientists and decision scientists—assume

  that the payoff function u(, x) is known: the decision maker is presumed to have a good sense of how actions and states combine to create outcomes.

  This assumption, however, is highly unrealistic. The credit card fraud ex-

  ample supplied by the authors is convincing. What is the long- term cost

  to a bank of approving a fraudulent transaction or labeling a legitimate

  transaction a suspected fraud?

  Organizations can spend resources to improve
both their prediction preci-

  sion and their judgment quality. Agrawal, Gans, and Goldfarb characterize

  the solution to this optimization problem. Their main result is that, under

  reasonable assumption, investment in prediction and investment in judg-

  ment are complementary (Proposition 2). Investing in prediction makes

  investment in judgment more benefi cial in expected value.

  This complementarity suggests that moving from a situation where

  prediction is prohibitively expensive to one where it is economical should

  increase the returns to judgment. In this perspective, the AI revolution will

  lead to an increase in the demand for judgment. However, judgment is an

  intrinsically diff erent problem—one that cannot be solved through the anal-

  ysis of big data.

  Let me suggest an example. 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. Leaving aside the

  censored observations problem arising from the fact that we only see the

  grades of successful applicants who decide to enroll, we can hope that AI

  can provide a fairly accurate prediction of an applicant’s future grades given

  his or her observable characteristics. 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. The latter is a highly complex

  object that depends on whether our alums become the kind of inspiring,

  successful, and ethical people that will add to the academic reputation and

  fi nancial sustainability of our university. There is likely to be a connection

  112 Andrea Prat

  between grades and this long- term goal, but we are not sure what it is. In

  this setting, Agrawal, Gans, and Goldfarb teach us an important lesson.

  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 institutions. These questions cannot be answered within

  AI, but rather with more theory- driven retrospective approaches or perhaps

  more qualitative methodologies.

  As an organizational economist, I am particularly interested in the impli-

  cations of Agrawal, Gans, and Goldfarb’s model for the study of organi-

  zations. First, this chapter highlights the importance of the dynamics of

  decision- making—a seriously underresearched topic. In a complex world,

  organizations are not going to immediately collect all the information they

  could possibly need about all possible contingencies they may face. Bolton

  and Faure- Grimaud (2009), a source of inspiration for Agrawal, Gans, and

  Goldfarb, model a decision maker who can “think ahead” about future states

  of the world in yet unrealized states of nature. They show that the typical

  decision maker does not want to think through a complete action plan, but

  rather focus on key short- and medium- term decisions. Agrawal, Gans, and

  Goldfarb show that Bolton and Faure- Grimaud’s ideas are highly relevant

  for understanding how organizations are likely to respond to changes in

  information technology.

  Second, Agrawal, Gans, and Goldfarb also speak to the organizational

  economics literature on mission. Dewatripont, Jewitt, and Tirole (1999)

  develop a model where organizational leaders are agents whose type is

  unknown, as in Holmstrom’s (1999) career concerns paradigm. Each agent

  is assigned a mission, a set of measured variables that are used to evaluate

  and reward the agent. Dewatripont, Jewitt, and Tirole identify a tension

  between selecting a simple one- dimensional mission that will provide the

  agent with a strong incentive to perform well or a “fuzzy” multidimensional

  mission that will dampen the agent’s incentive to work hard but will more

  closely mirror the true objective of the organization.

  This tension is also present in Agrawal, Gans, and Goldfarb’s world.

  Should we give the organization a mission that is close to a pure prediction

  problem, like admitting students who will get high grades? The pro is that

  it will be relatively easy to assess the leader’s performance. The con is that

  the outcome may be weakly related to the organization’s ultimate objective.

  Or should we give the organization a mission that also comprises the judg-

  ment problem, like furthering the long- term academic reputation of our

  university? This mission would be more representative of the organization’s

  ultimate objective, but may make it hard to assess our leaders and give them

  a weak incentive to adopt new prediction technologies. One possible lesson

  from Agrawal, Gans, and Goldfarb is that, as the cost of adopting AI goes

  down, the moral hazard problem connected with judgment becomes rela-

  Comment 113

  tively more important, thus militating in favor of incentive schemes that

  reward judgment rather than prediction.

  Third, Agrawal, Gans, and Goldfarb’s section on reliability touches on

  an important topic. Is it better to have a technology that returns accurate

  predictions with a low probability or less accurate predictions with a higher

  probability? The answer to this question depends on the available judgment

  technology. Better judgment technology increases the marginal benefi t of

  prediction accuracy rather than prediction frequency. More broadly, this

  type of analysis can guide the design of AI algorithms. Given the mapping

  between states, actions, and outcomes, and given the cost of various pre-

  diction technologies, what prediction technology should the organization

  select? A general analysis of this question may require using information

  theoretical concepts, introduced to economics by Sims (2003).

  Fourth, Agrawal, Gans, and Goldfarb show that economic theory can

  make important contributions to the debate over how AI will aff ect optimal

  organization. There is a related area where the interaction between econo-

  mists and computer scientists can be benefi cial. Artifi cial intelligence typi-

  cally assumes a stable fl ow of instances. When a bank develops an AI- based

  system to detect fraud, it assumes that the available data, which is used to

  build and test the detection algorithm, comes from the same data- generating

  process as future data on which the algorithm will be applied. However,

  the underlying data- generating process is not an exogenously given natural

  phenomenon: it is the output of a set of human beings who are pursuing

  their own goals, like maximizing the chance of getting their nonfraudulent

  application accepted or maximizing their chance of defrauding the bank.

  These sentient creatures will in the long term respond to the fraud- detection

  algorithm by modifying their application strategy, for instance, by providing

  diff erent information or by exerting eff ort to modify the reported variables.

  This means that the data- generating process will be subject to a structural

  change and that this change will be endogenous to the fraud- detection algo-

  rit
hm chosen by the bank. A similar phenomenon occurs in the university

  admission example discussed above: a whole consulting industry is devoted

  to understanding admissions criteria and advising applicants on how to

  maximize their success chances. A change in admissions practices is likely

  to be refl ected in the choices that high school students make.

  If the data- generating process is endogenous and depends on the predic-

  tion technology adopted by the organization, the judgment problem identi-

  fi ed by Agrawal, Gans, and Goldfarb becomes even more complex. The

  organization must evaluate how other agents will respond to changes in the

  prediction technology. As, by defi nition, no data is available about not yet

  realized data- generating processes, the only way to approach this problem

  is by estimating a structural model that allows other agents to respond to

  changes in our prediction technology.

  114 Andrea Prat

  In conclusion, Agrawal, Gans, and Goldfarb make a convincing case

  that the AI revolution should increase the benefi t of improving our judg-

  ment ability. They also provide us with a tractable yet powerful framework

  to understand the interaction between prediction and judgment. Future

  research should focus on further understanding the implications of improve-

  ments in prediction technology on the optimal structure of organizations.

  References

  Arrow, Kenneth. J. 1974. The Limits of Organization. New York: W. W. Norton.

  Bolton, P., and A. Faure- Grimaud. 2009. “Thinking Ahead: The Decision Problem.”

  Review of Economic Studies 76: 1205– 38.

  Dewatripont, Mathias, Ian Jewitt, and Jean Tirole. 1999. “The Economics of Career

  Concerns, Part II: Application to Missions and Accountability of Government

  Agencies.” Review of Economic Studies 66 (1): 199– 21.

  Holmstrom, Bengt. 1999. “Managerial Incentive Problems: A Dynamic Perspective.”

  Review of Economic Studies 66 (1): 169– 82.

  Milgrom, Paul, and John Roberts. 1990. “The Economics of Modern Manufac-

  turing: Technology, Strategy, and Organization.” American Economic Review

  June: 511– 28.

  Sims, Christopher. 2003. “Implications of Rational Inattention.” Journal of Mone-

  tary Economics 50 (3): 665– 90.

  4

  The Impact of Artifi cial

 

‹ Prev