by Ajay Agrawal
(RFF) feature through which sellers can set a rebate value for any item
they sell (cash back or store coupon) as a reward for a buyer’s feedback.
If a seller chooses this option, then Taobao guarantees that the rebate will
be transferred from the seller’s account to a buyer who leaves high- quality
feedback. Importantly, feedback quality only depends on how informative it
is, rather than whether the feedback is positive or negative. Taobao measures
the quality of feedback with a NLP algorithm that examines the comment’s
content and length and fi nds out whether key features of the item are men-
tioned. Hence, the marketplace manages the market for feedback by forcing
the seller to deposit at Taobao a certain amount for a chosen period, so that
funds are guaranteed for buyers who meet the rebate criterion, which itself
is determined by Taobao.6
6. According to a Taobao survey (published in March 2012), 64.8 percent of buyers believed that they will be more willing to buy items that have the RFF feature, and 84.2 percent of buyers believed that the RFF option will make them more likely to write detailed comments.
How AI and Machine Learning Can Impact Market Design 581
Taobao’s motivation behind the RFF mechanism was to promote more
informative feedback, but as Li, Tadelis, and Zhow (2016) noted, economic
theory off ers some insights into how the RFF feature can act as a potent
signaling mechanism that will further separate higher- from lower- quality
sellers and products. To see this, recall the literature launched by Nelson
(1970) who suggested that advertising acts as a signal of quality. According
to the theory, advertising—which is a form of burning money—acts as a
signal that attracts buyers who correctly believe that only high- quality sellers
will choose to advertise. Incentive compatibility is achieved through repeat
purchases: buyers who purchase and experience the products of advertisers
will return in the future only if the goods sold are of high enough quality.
The cost of advertising can be high enough to deter low- quality sellers from
being willing to spend the money and sell only once because those sellers
will not attract repeat customers, and still low enough to leave profi ts for
higher- quality sellers. Hence, ads act as signals that separate high- quality
sellers, and in turn attract buyers to their products.
As Li, Tadelis, and Zhow (2016) argue, the RFF mechanism plays a
similar signaling role as ads do. Assuming that consumers express their ex-
periences truthfully in written feedback, any consumer who buys a product
and is given incentives to leave feedback will leave positive feedback only
if the buying experience was satisfactory. Hence, a seller will off er RFF
incentives to buyers only if the seller expects to receive positive feedback,
and this will happen only if the seller will provide high quality. If a seller
knows that their goods and services are unsatisfactory, then paying for feed-
back will generate negative feedback that will harm the low- quality seller.
Equilibrium behavior then implies that RFF, as a signal of high quality, will
attract more buyers and result in more sales. The role of AI was precisely
to reward buyers for information, not for positive feedback.
Li, Tadelis, and Zhou (2016) proceeded to analyze data from the period
where the RFF mechanism was featured and confi rmed that fi rst, as ex-
pected, more feedback was left in response to the incentives provided by the
RFF feature. More important, the additional feedback did not exhibit any
biases, suggesting that the NLP algorithms used were able to create the kind
of screening needed to select informative feedback. Also, the predictions of
the simple signaling story were borne out in the data, suggesting that using
NLP to support a novel market for feedback did indeed solve both the free-
rider problem and the cold- start problem that can hamper the growth of
online marketplaces.
23.4 Using AI to Reduce Search Frictions
An important application of AI and machine learning in online market-
places is the way in which potential buyers engage with the site and proceed
to search for products or services. Search engines that power the search of
582 Paul R. Milgrom and Steven Tadelis
products online are based on a variety of AI algorithms that are trained
to maximize what the provider believes to be the right objective. Often this
boils down to conversion, under the belief that the sooner a consumer con-
verts a search to a purchase, the happier the consumer is both in the short
and the long run. The rationale is simply that search itself is a friction, and
hence, maximizing the successful conversion of search activity to a purchase
reduces this friction.
This is not inconsistent with economic theory that has modeled search
as an inevitable costly process that separates consumers from the products
they want. The canonical search models in economics either build on the
seminal work of Stigler (1961), who assumes that consumers sample a fi xed
number of stores and choose to buy the lowest priced item, or more often,
on the models of McCall (1970) and Mortensen (1970), who posit that
a model of sequential search is a better description of consumer search
behavior. In both modeling approaches consumers know exactly what they
wish to buy.
However, it turns out that unlike the simplistic models of search employed
in economic theory, where consumers know what they are looking for and
the activity of search is just a costly friction, in reality, people’s search behav-
ior is rich and varied. A recent paper by Blake, Nosko, and Tadelis (2016)
uses comprehensive data from eBay to shed light on the search process with
minimal modeling assumptions. Their data show that consumers search
signifi cantly more than other studies—which had limited access to search
behavior over time—have suggested.
Furthermore, search often proceeds from the vague to the specifi c. For
example, early in a search a user may use the query “watch,” then refi ne it to
“men’s watch,” and later add further qualifying words such as color, shape,
strap type, and more. This suggests that consumers often learn about their
own tastes, and what product characteristics exist, as part of the search
process. Indeed, Blake et al. (2016) show that the average number of terms
in the query rises over time, and the propensity to use the default- ranking
algorithm declines over time as users move to more focused searches like
price sorting.
These observations suggest that marketplaces and retailers alike could
design their online search algorithms to understand search intent so as to
better serve their consumers. If a consumer is in the earlier, exploratory
phases of the search process, then off ering some breadth will help the con-
sumer better learn their tastes and the options available in the market. But
when the consumer is driven to purchase something particular, off ering a
narrower set of products that match the consumer’s preferences would be
better. Hence, machine learning and AI can play an instrumental role in
> recognizing customer intent.
Artifi cial intelligence and machine learning cannot only help predict a
customer’s intent, but given the large heterogeneity on consumer tastes, AI
How AI and Machine Learning Can Impact Market Design 583
can help a marketplace or retailer better segment the many customers into
groups that can be better served with tailored information. Of course, the
idea of using AI for more refi ned customer segmentation, or even personal-
ized experiences, also raises concerns about price discrimination. For ex-
ample, in 2012 the Wall Street Journal reported that “Orbitz Worldwide Inc.
has found that people who use . . . Mac computers spend as much as 30%
more a night on hotels, so the online travel agency is starting to show them
diff erent, and sometimes costlier, travel options than Windows visitors see.
The Orbitz eff ort, which is in its early stages, demonstrates how tracking
people’s online activities can use even seemingly innocuous information—in
this case, the fact that customers are visiting Orbitz .com from a Mac—to
start predicting their tastes and spending habits.”7
Whether these practices of employing consumer data and AI will help or
harm consumers is not obvious, as it is well known from economic theory
that price discrimination can either increase or reduce consumer welfare. If,
on average, Mac users prefer staying at fancier and more expensive hotels
because owning a Mac is correlated with higher income and tastes for luxury,
then the Orbitz practice is benefi cial because it shows people what they want
to see and reduces search frictions. However, if this is just a way to extract
more surplus from consumers who are less price sensitive, but do not neces-
sarily care for the snazzier hotel rooms, then it harms these consumers.
There is currently a lot of interest in policy circles regarding the poten-
tial harms to consumers from AI- based price discrimination and market
segmentation. McSweeny and O’Dea (2017) suggest that once AI is used to
create more targeted market segments, this may not only have implications
only for consumer welfare, but for antitrust policy and market defi nitions for
mergers. But, as Gal and Elkin- Koren (2017) suggest, the same AI- targeting
tools used by retailers and marketplaces to better segment consumers may
be developed into tools for consumers that will help them shop for better
deals and limit the ways in which marketplaces and retailers can engage in
price discrimination.
23.5 Concluding
Remarks
In its early years, classical economic theory paid little attention to market
frictions and treated information and computation as free. That theory led
to conclusions about effi
ciency, competitive prices for most goods, and full
employment of valuable resources. To address the failures of that theory,
economists began to study models with search frictions, which predicted that
price competition would be attenuated, that some workers and resources
7. See “On Orbitz, Mac Users Steered to Pricier Hotels,” Dana Mattioli, The Wall Street Journal, Aug. 23, 2012. https:// www .wsj .com/ articles/ SB1000142405270230445860457748882
2667325882.
584 Paul R. Milgrom and Steven Tadelis
could remain unemployed, and that it could be costly to distinguish reliable
trading partners from others. They also built markets for complex resource-
allocation problems in which computations and some communications were
centralized, lifting the burden of coordination from individual market par-
ticipants.
With these as the key frictions in the traditional economy, AI holds enor-
mous potential to improve effi
ciency. In this chapter, we have described
some of the ways that AI can overcome computational barriers, reduce
search frictions, and distinguish reliable partners. These are among the most
important causes of ineffi
ciency in traditional economies, and there is no
longer any question that AI is helping to overcome them, with the promise
of widespread benefi ts for all of us. As Roth (2002) noted, market designers
“cannot work only with the simple conceptual models used for theoretical
insights into the general working of markets. Instead, market design calls
for an engineering approach.” Artifi cial intelligence has already proven to
be a valuable tool in the economist- as-engineer tool box.
References
Blake, T., C. Nosko, and S. Tadelis. 2016. “Returns to Consumer Search: Evidence
from eBay.” 17th ACM Conference on Electronic Commerce (EC 2016), 531– 45.
Filippas, A., J. J. Horton, and J. M. Golden. 2017. “Reputation in the Long- Run.”
Working paper, Stern School of Business, New York University.
Gal, M. S., and N. Elkin- Koren. 2017. “Algorithmic Consumers.” Harvard Journal
of Law & Technology 30:1– 45.
Hutter, F., H. Hoos, and K. Leyton- Brown. 2011. “Sequential Model- Based Opti-
mization for General Algorithm Confi guration.” Conference paper, International
Conference on Learning and Intelligent Optimization. https:// link.springer .com
/ chapter/ 10.1007/ 978-3-642-25566-3_40.
Leyton- Brown, K., P. R. Milgrom, and I. Segal. 2017. “Economics and Computer
Science of a Radio Spectrum Reallocation.” Proceedings of the National Academy
of Sciences 114 (28): 7202– 09. www .pnas .org/ cgi/ doi/ 10.1073/ pnas.1701997114.
Li, L. I., S. Tadelis, and X. Zhou. 2016. “Buying Reputation as a Signal of Qual-
ity: Evidence from an Online Marketplace.” NBER Working Paper no. 22584,
Cambridge, MA.
Masterov, D. V., U. F. Mayer, and S. Tadelis. 2015. “Canary in the E-commerce
Coal Mine: Detecting and Predicting Poor Experiences Using Buyer- to-Seller
Messages.” In Proceedings of the Sixteenth ACM Conference on Economics and
Computation, EC ‘15, 81– 93.
Mayzlin, D., Y. Dover, and J. Chevalier. 2014. “Promotional Reviews: An Empirical
Investigation of Online Review Manipulation.” American Economic Review 104
(8): 2421– 55.
McCall, J. J. 1970. “Economics of Information and Job Search.” Quarterly Journal
of Economics 84 (1): 113– 26.
McSweeny, T., and B. O’Dea. 2017. “The Implications of Algorithmic Pricing for
How AI and Machine Learning Can Impact Market Design 585
Coordinated Eff ects, Analysis and Price Discrimination Markets in Antitrust
Enforcement.” Antitrust 32 (1): 75– 81.
Milgrom, P. R. 2017. Discovering Prices: Auction Design in Markets with Complex
Constraints. New York: Columbia University Press.
Milgrom, P. R., D. C. North, and B. R. Weingast. 1990. “The Role of Institutions
in the Revival of Trade: The Law Merchant, Private Judges, and the Champagne
Fairs.” Economics and Politics 2 (1): 1– 23.
Milgrom, P. R., and I. Segal. 2017. “Deferred Acceptance Auctions and Radio Spec-
trum Reallocation.” Working paper.
Mortensen, D. T. 1970. “Job Search, the Duration of Unemployment and the Phil-
lips Curve.” American Economic Review 60 (5): 847– 62.
Myerson, R. B. 1981. “Optimal Auction Design.” Mathematics of Operations
Research 6 (1)
: 58– 73.
Nelson, P. 1970. “Information and Consumer Behavior.” Journal of Political
Economy 78 (2): 311– 29.
Newman, N., A. Fréchette, and K. Leyton- Brown. Forthcoming. “Deep Optimiza-
tion for Spectrum Repacking.” Communications of the ACM (CACM).
Nosko, C., and S. Tadelis. 2015. “The Limits of Reputation in Platform Markets:
An Empirical Analysis and Field Experiment.” NBER Working Paper no. 20830,
Cambridge, MA.
Ostrovsky, M., and M. Schwartz. 2017. “Reserve Prices in Internet Advertising: A
Field Experiment.” Working paper.
Stigler, G. J. 1961. “The Economics of Information.” Journal of Political Economy
69 (3): 213– 25.
Roth, A. E. 2002. “The Economist as Engineer: Game Theory, Experimentation,
and Computation as Tools for Design Economics.” Econometrica 70 (4): 1341– 78.
Roth, A. E. 2015. Who Gets What—and Why: The New Economics of Matchmaking
and Market Design. New York: Houghton Miffl
in Harcourt.
Vickrey, W. 1961. “Counterspeculation, Auctions, and Competitive Sealed Tenders.”
Journal of Finance, American Finance Association 16 (1): 8– 37.
Xu, H., D. Liu, H. Wang, and A. Stavrou. 2015. “E- commerce Reputation Manipu-
lation: The Emergence of Reputation- Escalation- as-a- Service.” Proceedings of
24th World Wide Web Conference (WWW 2015):1296– 306.
Zervas, G., D. Proserpio, and J. W. Byers. 2015. “A First Look at Online Reputa-
tion on Airbnb, Where Every Stay Is Above Average.” Working paper, Boston
University.
24
Artifi cial Intelligence and
Behavioral Economics
Colin F. Camerer
24.1 Introduction
This chapter describes three highly speculative ideas about how artifi cial
intelligence (AI) and behavioral economics may interact, particular in future
developments in the economy and in research frontiers. First note that I will
use the terms AI and machine learning (ML) interchangeably (although AI
is broader) because the examples I have in mind all involve ML and predic-
tion. A good introduction to ML for economists is Mullainathan and Spiess
(2017), and other chapters in this volume.
The fi rst idea is that ML can be used in the search for new “behavioral”-
type variables that aff ect choice. Two examples are given, from experimen-