by Ajay Agrawal
version of such a regulation with bipartisan support in September 2017 (the
SELF DRIVE Act, H.R. 3388).
From the perspective of innovation, a centralized AI regulatory system
presents a number of trade- off s. On the one hand, relative to tort laws that
examine liability cases ex post through judges and juries, ex ante regulations
and safety preemption would signifi cantly reduce the degree of uncertainty
regarding liability risk.9 Reduction in uncertainty, in general, increases R&D
and other complementary investment. Furthermore, harmonizing diff erent,
slow- moving state- wide regulations could also speed up experimentation
and adoption. In the case of autonomous vehicles, as of September 2017,
some testing was explicitly allowed in less than half of the states with dif-
ferent degrees of restriction and safety standards.
On the other hand, federal regulation could trade off certainty with fl exi-
bility. With the fast- changing landscape of AI technologies, federal agencies
may not have suffi
cient information in the early development stage to set
eff ective standards.10 If such regulations were hard to change, they could
infl uence the rate and direction of innovation in undesirable ways.
20.4.3 Allocation of Liability Risk across the Vertical Chain
Artifi cial intelligence and sophisticated robotics are often complex tech-
nologies that involve multiple suppliers of software and hardware and that
may require high degrees of integration between diff erent components.
Furthermore, AI technologies, like other general purpose technologies (such
as polymers), once developed for the fi rst few areas, may later be developed
for a wide variety of applications at a lower cost.
Current laws, such as component parts and sophisticated purchaser doc-
trines, stipulate that component suppliers are not liable unless the com-
ponent per se is defective or the process of integrating the component has
caused the adverse eff ect (Hubbard 2015). In practice, however, these laws
may be inadequate in certain circumstances and may expose component
8. Federal preemption of state laws may be explicit or implicit, with the former providing signifi cantly greater clarity. In the case of FDA preemption in Riegel v. Medtronic, Inc. (2008), the US Supreme Court ruled that manufacturers of FDA- approved devices that went through the pre- market approval process are protected from liability claims under state laws. In Wyeth v.
Levine (2009), however, the US Supreme Court ruled that Vermont tort law was not preempted.
9. Kaplow (1992) provides a general economic analysis of rules versus standards; that is, whether laws should be given content ex ante or ex post. The basic trade- off s depend on factors including the frequency and the degree of heterogeneity of adverse events, as well as the relative costs of individuals in learning and applying the law.
10. For example, for autonomous vehicles state regulators currently diff er in their opinions about whether cars without steering wheels or brake pedals should be allowed on public roads for testing and operation purposes.
Economics of Tort Liability and Innovation in Artifi cial Intelligence 501
suppliers to disproportionately high liability risk relative to their expected
revenue. Evidence from Galasso and Luo (2018) suggests that in such situa-
tions, downstream innovation may suff er. It would be interesting for future
research to examine how liability costs should be allocated across compo-
nent producers and its impacts on innovation, and under what conditions
policymakers should consider applying exemption regulations, such as the
BAAA enacted in the medical implants industry.
It would also be interesting to examine how liability rules, apart from
their direct impacts on innovation, infl uence fi rm boundaries, which, in
turn, could aff ect innovation. For example, rules such as the BAAA may
discourage vertical integration because its exemption applies only to com-
ponent material suppliers suffi
ciently removed from downstream activities.
Similarly, liability rules may also infl uence how products and services are
designed. For example, they may encourage more modular designs to better
insulate liability risk across diff erent components.
20.4.4 Liability Risk and Market Structure
Relatedly, it would be interesting to better understand the interplay
between liability risk and industry market structure and whether changes
in market structure driven by liability risk have long- term consequences for
innovation (Agrawal et al. 2014).
How liability risk aff ects fi rms of diff erent sizes is likely to depend on the
empirical context. Plaintiff s may be more likely to target larger and cash-
rich fi rms (Cohen, Gurun, and Kominers [2014] fi nd this pattern in patent
litigation cases). At the same time, larger fi rms are better at withstanding
high liability risk because they have greater resources both to self- insure and
to provide more generous indemnifi cation contracts to suppliers.
One may argue that liability insurance could insulate producers from
potential liability concerns. However, in the early stages of AI technolo-
gies the market for liability insurance may not be fully developed, or even
exist, due to insuffi
cient data on adverse events of a particular nature and
their damages. Even with well- developed insurance markets, high liability
risk may result in high premiums, which can be prohibitively expensive for
smaller fi rms and, thus, deter entry.
20.4.5 Liability
Litigation
An important feature of an eff ective liability system is that disputes are
resolved quickly. Longer settlement delays are typically associated with
higher transaction costs for the negotiating parties. More importantly,
delays and uncertainty in the process mean slower diff usion of the AI tech-
nology at the center of the dispute.
It is not obvious whether liability suits related to AI technologies will be
easier to settle than those involving other technologies. In particular, the
complexity of these new technologies and certain types of human- machine
502 Alberto Galasso and Hong Luo
interactions may reduce the litigants’ ability to fi nd a compromise. That said,
classic models of pretrial negotiations predict a higher likelihood of settle-
ment when information asymmetries between litigating parties are reduced
(Spier 2007). Manufacturers of AI technologies may fi nd it in their own
interest to design the machine’s data- recording capability in ways that facili-
tate the discovery process and speed up settlement. In cases where manu-
facturers lack such incentives, mandates of certain designs may be necessary
if they are clearly effi
ciency enhancing. Once again, how eff ective these data
capabilities of AI technologies are in facilitating dispute resolution would
also depend on the ability of the court system to understand and interpret
data, on private incentives for data sharing, and on whether policies are in
place to discourage misrepresentation and manipulation of data.
20.4.6 Liability Risk and Intellectual Property Protection
The likely impacts of liability risk on innovation would also depend on the
> strength of intellectual property (IP) rights. Intuitively, when IP rights are
strong, fi rms can invest in safer products and recover their investments by
charging a price premium. However, if competitors can easily copy and sell
these products or features, the incentive to innovate in the fi rst place would
decrease. The above considerations may be diff erent, however, if consumers
cannot easily distinguish between safer and less safe products, and their fears
about dangerous products suppress their demand for the entire product cate-
gory. For example, Jarrell and Pelzman (1985) show that one fi rm’s product
recalls may have negative reputational impacts on competitors and produc-
ers of related products. When such negative spillover is strong, fi rms with
other means of extracting rents (e.g., larger fi rms) may have the incentive to
invest in safety features and share them with fi rms in the industry so as to
maintain consumer demand for the whole industry.
Finally, in a cumulative innovation environment, as Green and Scotchmer
(1995) have shown, the allocation of IP rights among sequential innovators
may have important eff ects on their respective innovation incentives. Related
trade- off s are likely to also emerge for the allocation of liability damages
among sequential innovators.
20.5 Conclusion
This chapter has examined some of the basic economic trade- off s linking
liability risk with innovation incentives and the direction of technological
progress in the context of artifi cial intelligence and sophisticated robots.
Features of the liability system, such as the allocation of risk between pro-
ducers and consumers and the level of centralization in regulation, may
have a signifi cant impact on the development and diff usion of these new
technologies, as well as on the products and services that apply them. The
Economics of Tort Liability and Innovation in Artifi cial Intelligence 503
extent of these eff ects is likely to also depend on the market structure and
the organization of the vertical chain of innovation.
More broadly, our analysis supports the idea that the liability system and
its reforms can aff ect the rate and the direction of technological change, indi-
cating that these policies have dynamic eff ects on innovation incentives that
go beyond their short- term impact on the safety of the users and others. As
Finkelstein (2004) stresses, recognizing and estimating these dynamic eff ects
is crucial to evaluating the costs and benefi ts of policy reforms.
References
Aghion, P., A. Dechezleprêtre, D. Hemous, R. Martin, and J. Van Reenen. 2016.
“Carbon Taxes, Path Dependency, and Directed Technical Change: Evidence from
the Auto Industry.” Journal of Political Economy 124:1– 51.
Agrawal, A., I. Cockburn, A. Galasso, and A. Oettl. 2014. “Why Are Some Regions
More Innovative Than Others? The Role of Small Firms in the Presence of Large
Labs.” Journal of Urban Economics 81:149– 65.
Aronoff , M. 1995. “Market Study: Biomaterials Supply for Permanent Medical
Implants.” Journal of Biomaterials Applications 9:205– 60.
Cohen, L., U. Gurun, and S. D. Kominers. 2014. “Patent Trolls: Evidence from
Targeted Firms.” NBER Working Paper no. 20322, Cambridge, MA.
Del Rossi, A., and K. Viscusi. 2010. “The Changing Landscape of Blockbuster
Punitive Damages Awards.” American Law and Economics Review 12:116– 61.
European Parliament. 2017. Civil Law Rules on Robotics 2015/ 2103(INL)—16/ 02/
2017.
Feder, Barnaby. 1994. “Implant Industry Is Facing Cutback by Top Suppliers.” New
York Times, Apr. 25. https:// www .nytimes .com/ 1994/ 04/ 25/ us/ implant- industry
- is- facing- cutback- by- top- suppliers .html.
Finkelstein, A. 2004. “Static and Dynamic Eff ects of Health Policy: Evidence from
the Vaccine Industry.” Quarterly Journal of Economics 119:527– 64.
Galasso, A., and H. Luo. 2017. “Tort Reform and Innovation.” Journal of Law and
Economics 60:385– 412.
———. 2018. “How Does Product Liability Risk Aff ect Innovation? Evidence from
Medical Implants.” Working paper, Harvard Business School.
Green, J., and S. Scotchmer. 1995. “On the Division of Profi t in Sequential Innova-
tion.” RAND Journal of Economics 26:20– 33.
Hay, B. and K. Spier. 2005. “Manufacturer Liability for Harms Caused by Consum-
ers to Others.” American Economic Review 95:1700– 11.
House of Representatives. 1997. Subcommittee on Commercial And Administra-
tive Law.
Hubbard, F. 2015. “Allocating the Risk of Physical Injury from ‘Sophisticated
Robots’: Effi
ciency, Fairness, and Innovation.” In Robot Law, edited by R. Calo,
A. Froomkin, and I. Kerr. Cheltenham, UK: Edward Elgar Publishing.
Huber, P. 1989. Liability: The Legal Revolution and Its Consequences. New York: Basic Books.
Huber, P., and R. Litan, eds. 1991. The Liability Maze: The Impact of Liability Law on Safety and Innovation. Washington, DC: Brookings Institution.
504 Alberto Galasso and Hong Luo
Jarrell, G., and S. Peltzman. 1985. “The Impact of Product Recalls on the Wealth of
Sellers.” Journal of Political Economy 93:512– 36.
Kaplow, L. 1992. “Rules versus Standards: An Economic Analysis.” Duke Law Jour-
nal 42:557– 629.
Kerouac, J. 2001. “A Critical Analysis of the Biomaterials Access Assurance Act
of 1998 as Federal Tort Reform Policy.” Boston University Journal of Science and
Technology Law 7:327– 45.
Marchant, G., and R. Lindor. 2012. “The Coming Collision between Autonomous
Vehicles and the Liability System.” Santa Clara Law Review 52:1321– 40.
O’Reilly, B. 2014. “Patents Running Out: Time to Take Stock of Robotic Surgery.”
International Urogynecology Journal 25:711– 13.
Parchomovsky, G., and A. Stein. 2008. “Torts and Innovation.” Michigan Law
Review 107:285– 315.
Polinsky, M., and S. Shavell. 2010. “The Uneasy Case for Product Liability.” Harvard Law Review 123:1437– 92.
Porter, M. 1990. The Competitive Advantage of Nations. New York: Free Press.
Schwartz, A. 1988. “Products Liability Reform: A Theoretical Synthesis.” Yale Law
Journal 97:353– 419.
Schmucki, R. 1999. “Final Status Report on History of TMJ Litigation.” DuPont,
unpublished communication.
Scotchmer, S. 1999. “On the Optimality of the Patent Renewal System.” RAND
Journal of Economics 30:181– 96.
Spier, K. 2007. “Litigation.” In Handbook of Law and Economics, edited by A. M.
Polinsky and S. Shavell. Amsterdam: North- Holland.
Viscusi, K., and M. Moore. 1993. “Product Liability, Research and Development,
and Innovation.” Journal of Political Economy 101:161– 84.
Von Hippel, E. 2005. Democratizing Innovation. Cambridge, MA: MIT Press.
IV
Machine Learning and Economics
21
The Impact of Machine
Learning on Economics
Susan Athey
21.1 Introduction
I believe that machine learning (ML) will have a dramatic impact on the
fi eld of economics within a short time frame. Indeed, the impact of ML on
; economics is already well underway, and so it is perhaps not too diffi
cult to
predict some of the eff ects.
The chapter begins by stating the defi nition of ML that I will use in this
chapter, describing its strengths and weaknesses, and contrasting ML with
traditional econometrics tools for causal inference, which is a primary focus
of the empirical economics literature. Next, I review some applications of
ML in economics where ML can be used off the shelf: the use case in eco-
nomics is essentially the same use case that the ML tools were designed
and optimized for. I then review “prediction policy” problems (Kleinberg
et al. 2015), where prediction tools have been embedded in the context of
economic decision- making. Then, I provide an overview of the questions
considered and early themes of the emerging literature in econometrics and
statistics combining machine learning and causal inference, a literature that
is providing insights and theoretical results that are novel from the per-
spective of both ML and statistics/ econometrics. Finally, I step back and
Susan Athey is the Economics of Technology Professor at Stanford University Graduate
School of Business and a research associate of the National Bureau of Economic Research.
I am grateful to David Blei, Guido Imbens, Denis Nekipelov, Francisco Ruiz, and Stefan
Wager, with whom I have collaborated on many projects at the intersection of machine learning and econometrics and who have shaped my thinking, as well as to Mike Luca, Sendhil Mullainathan, and Hal Varian, who have also contributed to my thinking through their writing, lecture notes, and many conversations. 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/ c14009.ack.
507
508 Susan Athey
describe the implications of the fi eld of economics as a whole. Throughout,
I make reference to the literature broadly, but do not attempt to conduct a
comprehensive survey or reference every application in economics.
The chapter highlights several themes.
A fi rst theme is that ML does not add much to questions about identifi ca-
tion, which concerns when the object of interest, for example, a causal eff ect,
can be estimated with infi nite data, but rather yields great improvements