The Economics of Artificial Intelligence
Page 72
eff ect we measure does not appear to be infl uenced by the status of women
in that country. Instead, we present evidence that the algorithm is reacting
to spillovers across advertisers. Women are a prized demographic among
advertisers, both because they are often more profi table and because they
control much of the household expenditure. Therefore, profi t- maximizing
fi rms pay more to show ads to female eyeballs than male eyeballs, especially
in younger demographics. These spillovers across advertisers and the algo-
rithms’ attempts to cost- minimize given these spillovers explain the eff ect
we measure. Women are less likely to see an intended gender- neutral ad due
to crowding out eff ects.
To put it simply, our results are the result of these factors:
1. The ad algorithm is designed to minimize cost so that advertisers’ adver-
tising dollars will stretch further.
2. Other advertisers consider female eyeballs to be more desirable and
deliver a higher return on investment and therefore are willing to pay more
to have their ads shown to women than men.
Lambrecht and Tucker (forthcoming) explore apparent algorithmic bias,
which is the consequence of clear economic spillovers between the value of
a pair of eyeballs for one organization compared to another. Beyond ensur-
ing that, for example, fi rms advertising for jobs are aware of the potential
consequences, it is diffi
cult to know what policy intervention is needed or
the extent to which this should be thought of as a privacy issue rather than
analyzed through the already established policy tools set up to address dis-
crimination.
This kind of spillover, though, is another example of how in an intercon-
nected economy, models of privacy that stipulate privacy as an exchange
between a single fi rm and a single consumer may no longer be appropriate
for a connected economy. Instead, the way any piece of data may be used
by a single fi rm may itself be subject to spillovers from other entities in the
economy, again in ways that may not be easily foreseen at the time of data
creation.
17.6 Implications and Future Research Agenda
This chapter is a short introduction into the relationship between artifi cial
intelligence and the economics of privacy. It has emphasized three themes:
data persistence, data repurposing, and data spillovers. These three areas
may present some new challenges for the traditional treatment of privacy
within an individual’s utility function as they suggest challenges for the ways
we model how an individual may make choices about the creation of per-
Privacy, Algorithms, and Artifi cial Intelligence 435
sonal data that can later be used to inform an algorithm. At the highest level,
this suggests that future work on privacy in economics may focus on the
dynamics of privacy considerations amid data persistence and repurposing,
and the spillovers that undermine the clarity of property rights over data,
rather than the more traditional atomistic and static focus of our economic
models of privacy.
17.6.1 Future Research Agenda
To conclude this chapter, I highlight specifi c research questions that fall
under these three areas:
• Data Persistence
1. What causes consumers’ privacy preferences to evolve over time? How
stable are these preferences and for how long?
2. Are consumers able to correctly predict the evolution of their privacy
preferences as they get older?
3. Would regulations designed to restrict the length of time that compa-
nies can store data be welfare enhancing or reducing?
4. What infl uences the persistence of the value of data over the long run?
Are there some types of data that lose their value to algorithms quickly?
• Data Reuse
1. Do consumers appreciate the extent to which their data can be reused
and are they able to predict what their data may be able to predict?
2. What kind of regulations restricting data reuse may be optimal?
3. Do approaches to data contracting based on the blockchain or other
transaction cost- reducing technologies enable suffi
ciently broad contracts
(and the establishment of property rights) over data?
4. Are there any categories of data where reuse by algorithms should be
explicitly restricted?
• Data Spillovers
1. Are there any mechanisms (either theoretical or practical) that could be
used to ensure that people internalized the consequences of their creation
of data for others?
2. What is the best mechanism by which individuals may be able to assert
their right to exclusion from some types of data that are being broadly col-
lected (genetic data, visual data, surveillance data, etc.)?
3. Is there any evidence for the hypothesis of biased AI programmers,
leading to biased AI algorithms? Would eff orts to improve diversity in the
technology community reduce the potential for bias?
4. How much more biased are algorithms that appear to engage in data-
based discrimination than the counterfactual human process?
436 Catherine Tucker
References
Acquisti, A. 2010. “From the Economics to the Behavioral Economics of Privacy:
A Note.” In Ethics and Policy of Biometrics, edited by A. Kumar and D. Zhang,
23– 26. Lecture Notes in Computer Science, vol. 6005. Berlin: Springer.
Acquisti, A., C. R. Taylor, and L. Wagman. 2016. “The Economics of Privacy.”
Journal of Economic Literature, 52 (2): 442–92.
Adee, S. 2015. “Can Data Ever Be Deleted? New Scientist 227 (3032): 17.
Agrawal, A., J. Gans, and A. Goldfarb. 2016. “The Simple Economics of Machine
Intelligence.” Harvard Business Review, Nov. 17. https:// hbr .org/ 2016/ 11/ the
- simple- economics- of-machine- intelligence.
Athey, S., C. Catalini, and C. Tucker. 2017. “The Digital Privacy Paradox: Small
Money, Small Costs, Small Talk.” Technical Report, National Bureau of Eco-
nomic Research.
Bertocchi, G., and A. Dimico. 2014. “Slavery, Education, and Inequality.” European
Economic Review 70:197– 209.
Chiou, L., and C. E. Tucker. 2014. “Search Engines and Data Retention: Implica-
tions for Privacy and Antitrust.” MIT Sloan Research Paper no. 5094-14, Massa-
chusetts Institute of Technology.
Custers, B., T. Calders, B. Schermer, and T. Zarsky. 2012. “Discrimination and Pri-
vacy in the Information Society.” In Volume 3 of Studies in Applied Philosophy,
Epistemology and Rational Ethics Berlin: Springer.
Datta, A., M. C. Tschantz, and A. Datta. 2015. “Automated Experiments on Ad Pri-
vacy Settings.” Proceedings on Privacy Enhancing Technologies 2015 (1): 92– 112.
Dietvorst, B. J., J. P. Simmons, and C. Massey. 2016. “Overcoming Algorithm Aver-
sion: People Will Use Imperfect Algorithms If They Can (Even Slightly) Modify
Them.” Management Science https:// doi .org/ 10.1287/ mnsc.2016.2643.
Farrell, J. 2012. “Can Privacy Be Just Another Good?” Journal on Telecommunica-
tions and High Technology Law 10:251.
Goldfarb, A., and C. Tucker. 2012. “Shifts in Privacy Conc
erns.” American Economic
Review: Papers and Proceedings 102 (3): 349– 53.
Kosinski, M., D. Stillwell, and T. Graepel. 2013. “Private Traits and Attributes are
Predictable from Digital Records of Human Behavior.” Proceedings of the Na-
tional Academy of Sciences 110 (15): 5802– 05.
Lambrecht, A., and C. Tucker. Forthcoming. “Algorithmic Discrimination? Appar-
ent Algorithmic Bias in the Serving of Stem Ads.” Management Science.
———. 2013. “When Does Retargeting Work? Information Specifi city in Online
Advertising.” Journal of Marketing Research 50 (5): 561– 76.
———. 2017. “Can Big Data Protect a Firm from Competition?” CPI Antitrust
Chronicle, Jan. 2017. https:// www .competitionpolicyinternational .com/ can- big
- data- protect- a- fi rm- from- competition/.
Marthews, A., and C. Tucker. 2014. “Government Surveillance and Internet Search
Behavior.” Unpublished manuscript, Massachusetts Institute of Technology.
McDonald, A. M., and L. F. Cranor. 2008. “The Cost of Reading Privacy Policies.”
Journal of Law and Policy for the Information Society 4 (3): 543– 68.
Miller, A., and C. Tucker. 2017. “Privacy Protection, Personalized Medicine and
Genetic Testing.” Management Science. https:// doi .org/ 10.1287/ mnsc.2017.2858.
———. 2018. “Historic Patterns of Racial Oppression and Algorithms.” Unpub-
lished manuscript, Massachsetts Institute of Technology.
O’Neil, C. 2017. Weapons of Math Destruction: How Big Data Increases Inequality
and Threatens Democracy. Portland, OR: Broadway Books.
Rubinstein, A. 2006. “Discussion of ‘Behavioral Economics.’ ” Unpublished manu-
Privacy, Algorithms, and Artifi cial Intelligence 437
script, School of Economics, Tel Aviv University, and Department of Economics,
New York University.
Sokoloff , K. L., and S. L. Engerman. 2000. “Institutions, Factor Endowments, and
Paths of Development in the New World.” Journal of Economic Perspectives 14
(3): 217– 32.
Strotz, R. H. 1955. “Myopia and Inconsistency in Dynamic Utility Maximization.”
Review of Economic Studies 23 (3): 165– 80.
Sweeney, L. 2013. “Discrimination in Online Ad Delivery.” ACM Queue 11 (3): 10.
Tucker, C. 2014. “Social Networks, Personalized Advertising, and Privacy Controls.”
Journal of Marketing Research 51 (5): 546– 62.
Varian, H. R. 1996. “Economic Aspects of Personal Privacy.” Working paper, Uni-
versity of California, Berkeley.
Warren, S. D., and L. D. Brandeis. 1890. “The Right to Privacy.” Harvard Law Review 4 (5): 193– 220.
18
Artifi cial Intelligence and
Consumer Privacy
Ginger Zhe Jin
Thanks to big data, artifi cial intelligence (AI) has spurred exciting innova-
tions. In the meantime, AI and big data are reshaping the risk in consumer
privacy and data security. In this chapter, I fi rst defi ne the nature of the
problem and then present a few facts about the ongoing risk. The bulk of
the chapter describes how the US market copes with the risk in current
policy environment. It concludes with key challenges facing researchers and
policymakers.
18.1 Nature of the Problem
In early 1980s, economists tended to think of consumer privacy as an
information asymmetry within a focal transaction: for example, consumers
want to hide their willingness to pay just as fi rms want to hide their real
marginal cost, and buyers with less favorable information (say a low credit
score) prefer to withhold it just as sellers want to conceal poor product qual-
ity (Posner 1981; Stigler 1980). Information economics suggests that both
buyers and sellers have an incentive to hide or reveal private information,
and these incentives are crucial for market effi
ciency. In the context of a
single transaction, less privacy is not necessarily bad for economic effi
ciency.
Data technology that reveals consumer type could facilitate a better match
Ginger Zhe Jin is professor of economics at the University of Maryland and a research associate of the National Bureau of Economic Research.
I am grateful to Ajay Agrawal, Joshua Gans, and Avi Goldfarb for inviting me to contribute to the 2017 NBER Conference on the Economics of Artifi cial Intelligence, and to Catherine Tucker, Andrew Stivers, and conference participants for inspiring discussion and comments.
All errors are mine. 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
/ c14034.ack.
439
440 Ginger Zhe Jin
between product and consumer type, and data technology that helps buyers
to assess product quality could encourage high- quality production.
New concerns arise because technological advances, which have enabled
radical decline in the cost of collecting, storing, processing, and using data
in mass quantities, extend information asymmetry far beyond a single trans-
action. These advances are often summarized by the terms “big data” and
“AI.” By big data, I mean large volume of transaction- level data that could
identify individual consumers by itself or in combination with other data
sets. The most popular AI algorithms take big data as an input in order to
understand, predict, and infl uence consumer behavior. Modern AI, used
by legitimate companies, could improve management effi
ciency, motivate
innovations, and better match demand and supply. But AI in the wrong
hands also allows the mass production of fraud and deception.
Since data can be stored, traded, and used long after the transaction,
future data use is likely to grow with data processing technology such as
AI. More important, future data use is obscure to both sides of the transac-
tion when the buyer decides whether to give away personal data in a focal
transaction. The seller may be reluctant to restrict data use to a particular
purpose, a particular data- processing method or a particular time horizon
in light of future data technology. Even if it does not plan to use any data
technology itself, it can always sell the data to those that will use it. These
data markets motivate the seller to collect as much information as consum-
ers are willing to give.
Sophisticated consumers may anticipate the uncertainty and hesitate to
give away personal data. However, in many situations, identity and payment
information are crucial (or made crucial) to complete the focal transac-
tion, leaving even the most sophisticated consumers to trade off between
immediate gains from the focal transaction and potential loss from future
data use. One may argue that future data use is simply a new attribute of
the product traded in the focal transaction; as long as the attribute is clearly
conveyed between buyer and seller (say via a well- written privacy policy),
sellers in a competitive market will respect buyer preference for limited data
use. Unfortunately, this attribute is not currently well defi ned at the time of
the focal transaction, and it can evolve over time in ways that depend on the
seller’s data policy but are completely out of the buy
er’s view, control, ability
to predict, or ability to value. This ongoing information asymmetry, if not
addressed, could lead to a lemon’s market (with respect to future data use).
Incomplete information about future data use is not the only problem
lurking in the interaction between AI and consumer privacy. There are at
least two other problems related to the uncertainty about future data use
and value: one is externality and the other is commitment.
To be clear, future data use can be benefi cial or detrimental to consumers,
thus rational consumers may prefer to share personal data to some extent
(Varian 1997). However, benefi ts from future data use—for example, bet-
Artifi cial Intelligence and Consumer Privacy 441
ter consumer classifi cation, better demand prediction, or better product
design—can usually be internalized by the collector of the information via
internal data use or through the sale of data to third parties. In contrast,
damages from future misuse—for example, identity theft, blackmail, or
fraud—often accrue not to the collector but to the consumer. Because it
is often hard to trace back consumer harm to a particular data collector,
these damages may not be internalized by either the data collector or by
consumers in their choices about how to interact with the collector. This is
partly because the victim consumer may have shared the same information
with hundreds of sellers, and she has no control over how each piece of
information may get into the wrong hands. The asymmetry between accru-
able benefi ts and nonaccountable damages amounts to negative externality
from sellers to buyers.1 If there is no way to track back to the origin, sellers
have an incentive to overcollect buyer information.2
This diffi
culty in tracing damages back to actions by the data collec-
tor, together with uncertainty about future use and ongoing information
asymmetry about collector practices, also triggers a commitment problem.
Assuming consumers care about data use, every seller has an incentive to
boast about having the most consumer- friendly data policy in the focal
transaction, but will also retain the option to renege after data collection.
There might be some room to enforce declared data policy- specifi c promises,
if the seller’s actual practice is revealed to the public and found to contradict