The Economics of Artificial Intelligence
Page 83
i
i
incorporating technology i into the fi rm’s product, and r is the probability i
2. It is important to note that this chapter focuses solely on the likely impacts on innovation and the direction of technological change. We refer interested readers to Hay and Spier (2005) and Polinsky and Shavell (2010) for an overall welfare discussion of the liability system and its features, and to Marchant and Lindor (2012) and Hubbard (2015) and the references within for details of tort law and an exploration of their applications to autonomous vehicles and sophisticated robots.
Economics of Tort Liability and Innovation in Artifi cial Intelligence 495
that the use of the product will result in personal injuries. The expected lia-
bility cost given that injury happens is H , which captures the (conditional)
probability that a liability suit will be fi led and the expected cost that the
fi rm will face if involved in such a suit. We expect H to be positive even if the fi rm is fully insured against claims for monetary damages because liability suits also invariably result in opportunity costs of employee time and
fi rm resources, as well as in reputational damage.
The fi rm’s expected profi t, net of liability risk from selling a product incor-
porating technology i, is
= b r H.
i
i
i
We denote the technology that the fi rm currently uses as O and consider
the fi rm’s decision to develop a new technology, which we denote as N. We
assume a simple R&D process such that successful development takes place
with probability p(x) = x if the innovator incurs a research cost C(x) = x 2 / 2.
As in Aghion et al. (2016), we refer to x as the “innovation intensity,” which
captures the likelihood of successfully developing a new technology. In this
setting, the problem for the innovating fi rm is
x 2
max x
+ (1 x)
N
O
x
2
which yields the following:
(1)
x* =
= b
b + ( r
r ) H.
N
O
N
O
O
N
Formula (1) provides some basic insights into the relationship between
liability and innovation. First, at the intensive margin, the sign of the deriva-
tive of x* with respect to H captures the directional eff ect of an increase in liability risk on innovation intensity. Thus, an increase in liability risk
suppresses innovation incentives for new technologies that are riskier than
the current technology ( r > r ) but encourages new technologies that are N
O
safer ( r < r ). In other words, changes in liability risk aff ect the type of N
O
technologies in which a fi rm invests and infl uence the direction of innova-
tion. Second, investment in innovation takes place if the profi t potential of
the new technology is greater than its liability risk, relative to the old tech-
nology—that is, x* > 0 only if b – b > ( r – r ) H. Thus, at the extensive N
O
N
O
margin, marginal changes in liability risk will not aff ect whether the fi rm
develops the new technology if it is expected to be highly profi table relative
to the existing one (i.e., b – b is very large) unless it is extremely risky. In N
O
contrast, liability concerns will matter more for technologies “at the margin”
(i.e., the improvement in expected profi tability is modest).
Galasso and Luo (2017) extend this stylized model to the medical setting,
in which physicians (i.e., the direct users of technologies) face malpractice
liability risk. Changes in their liability exposure aff ect innovation incentives
in medical technologies through the demand channel. Assuming that ideas
496 Alberto Galasso and Hong Luo
for new technologies ( b , r ) are random draws from a bivariate distribution N
N
(as in Scotchmer 1999), they show that the overall eff ect of tort reforms that
reduce physicians’ liability risk on innovation incentives is ambiguous and
depends on the characteristics of the existing technology ( b – r ).
N
O
The main message of this illustrative model is that the link between lia-
bility and innovation is more complex and nuanced than the simple view of
“liability chills innovation,” which ignores the potential encouraging eff ect
of liability risk on a potentially broad set of innovations that help fi rms and
their customers manage risk.
20.3 Empirical Evidence on Liability and Innovation
In a pioneering study, Viscusi and Moore (1993) examine the relationship
between product- liability insurance costs and fi rms’ research and develop-
ment (R&D) investments, using a data set covering large US manufacturing
fi rms in multiple industries between 1980 and 1984. They document a sig-
nifi cant positive correlation between the expected liability insurance costs
and fi rms’ R&D intensity when such costs are low or moderate. Only when
liability costs are very high the correlation is negative. Furthermore, the
liability- innovation link is driven mainly by product rather than by process
R&D. They interpret these results as evidence that, on average, product
liability, rather than discouraging innovation, promotes fi rm investment in
product safety (likely through product design).
Galasso and Luo (2017) examine whether tort reforms that reduce physi-
cians’ liability exposure to medical malpractice litigation aff ect incentives
to develop new medical technologies. Diff erent from the focus on product
liability in Viscusi and Moore (2017), they examine how liability costs that
users (physicians) face aff ect upstream research investment. It is worth not-
ing that such a perspective broadens the scope of innovation from product
safety design to include a wide variety of complementary technologies that
help physicians manage risk, such as monitoring and diagnostic devices
and devices used in complex procedures to reduce the likelihood of adverse
events. Because these technologies are not themselves subject to product lia-
bility claims, they are more likely to be infl uenced by changes in user liability
through the demand channel than by product liability.
Using a panel data set for the period of 1985– 2005, Galasso and Luo
(2017) fi nd that, on average, the introduction of noneconomic damage caps
in a state is associated with a 15 percent reduction in medical device pat-
enting. The eff ect is, however, highly heterogeneous: tort reforms have the
largest negative impact in medical fi elds in which the probability of a mal-
practice claim is the highest, and they do not seem to aff ect patenting of the
highest or the lowest quality. These results are consistent with the idea that
the decline in innovation is driven primarily by the reduced demand from
Economics of Tort Liability and Innovation in Artifi cial Intelligence 497
physicians for safer technologies or complementary technologies that help
them manage risk. The welfare loss from such a large decline in quantity,
however, appears not as worrying because patents wi
th the highest impacts
are not negatively aff ected.
Galasso and Luo (2018) study the medical implant industry in the early
1990s, during which the liability risk faced by raw material suppliers signifi -
cantly increased relative to the risk faced by downstream producers. Vitek
was a leading producer of jaw (temporomandibular joint) implants in the
1980s. Its Food and Drug Administration (FDA)- approved products were
considered state of the art and safe for use by oral surgeons across the
United States (Schmucki 1999). In the late 1980s, unexpected and wide-
spread problems arose with Vitek’s products. Vitek fi led for bankruptcy
in 1990 under a deluge of lawsuits. Following Vitek’s bankruptcy, implant
patients started to fi le a large number of lawsuits against DuPont, a raw
material supplier for Vitek’s implants and a large fi rm with “deep pockets.”3
The consensus among industry observers is that these events generated
a substantial increase in the perceived liability risk faced by fi rms that sup-
plied materials to producers of permanent implants, many of which had
withdrawn from this market. This view is well summarized in a 1994 report
on the status of the biomaterial market (Aronoff 1995), which links this fear
of product liability suits to the jaw implant litigation. Eventually, DuPont
won all the lawsuits, but the process took ten years and cost over $40 million
(House of Representatives 1997). In contrast, DuPont’s revenue from these
implants totaled only a few thousand dollars.
Galasso and Luo (2018) compare the rate of patenting in implant
devices—excluding technologies involved in these litigations—to patenting
in a control group of nonimplant medical technologies whose suppliers were
not aff ected by the heightened litigation risk. The diff erence- in-diff erences
(DID) results show, overall, a substantial decrease in the number of new
patents for implants in the fi ve years after Vitek’s bankruptcy in 1990. Time-
specifi c eff ects show that implant and nonimplant technologies exhibited
parallel increasing trends before 1990, and that the negative eff ect on implant
technologies was immediate after 1990 and increased in magnitude over
time. The signifi cant drop in innovation in medical implants appears to have
been largely driven by device producers’ expectation of a supply shortage
of material inputs.
To address this problem, in 1998 the US Congress passed the Biomaterial
Access Assurance Act (BAAA), which exempts biomaterial suppliers for
medical implants from liabilities as long as they do not engage in the design,
3. In parallel, problems also surfaced with silicone breast implants. Also in this case, a leading implant manufacturer fi led for bankruptcy, and silicone suppliers were named as defendants in numerous lawsuits (Feder 1994).
498 Alberto Galasso and Hong Luo
production, testing, and distribution of the implants. The BAAA is one of
the few federal liability reforms, an area of legislation typically reserved for
the states (Kerouac 2001).4
Together, the empirical evidence in Viscusi and Moore (1993) and Galasso
and Luo (2017, 2018) challenges the simple view that “liability chills innova-
tion.” All three papers suggest that the link between liability and innovation
depends on the context, including the nature of the innovation, the level
of the liability risk, and the value of the technology. Furthermore, liability
risk aff ecting one area may impact innovation incentives in other, vertically
related segments. More research is needed to understand the complex and
nuanced links between liability and innovation, and whether targeted poli-
cies can address these issues.
20.4 Tort Liability and the Development of AI Technologies
The liability system may aff ect innovation incentives of AI technologies
and sophisticated robots in multiple ways, and the development of these
technologies may, in turn, demand adjustments to the law. Below, we focus
on a number of areas and highlight some of the economic trade- off s that
deserve further examination, both theoretically and empirically.5
20.4.1 Allocation of Liability Risk between Producers and Consumers
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. Eff ective policies would
require a basic understanding of the relationship between humans and AI
technologies—for example, whether they are substitutes or complements
(Agrawal, Gans, and Goldfarb, chapter 3, this volume ).
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 individual users to control risk. For example, the
operator of a fl eet of self- driving cars would have the data and predictive
capability to provide instantaneous warnings of an adverse event. The cost
4. Examples of such federal policies include the General Aviation Revitalization Act of 1994, which exempts makers of small aircraft from liability for planes after eighteen years, and the National Childhood Vaccine Injury Act of 1986, which limits liability for drug companies and creates a no- fault compensation system for those injured by vaccines.
5. It is important to note that the likely impacts on innovation incentives would also depend on fi rms’ ability to write contracts and the development of the insurance markets (Schwartz 1988). We leave the discussion of these important topics and the interplay between liability law and contract law for future work. In situations in which externality (harm to third parties) is high and in the early stage of AI technologies, during which the insurance market may not be well developed or even exist, the roles of these systems are likely to be more limited than for mature technologies.
Economics of Tort Liability and Innovation in Artifi cial Intelligence 499
of observing systematic, hazardous user behaviors may also become suffi
-
ciently low such that it would be more effi
cient for producers to take precau-
tions through product redesign. How such a shift might aff ect innovation
incentives would depend on how producer liability is specifi ed, especially
whether the long- term social benefi ts are included in the analysis of the
producer’s liability.
On the other hand, during the transitional period of an AI technology,
substantial human supervision may still be required. Such interaction be-
tween AI and humans may not be obvious and diffi
cult to predict. For ex-
ample, it may actually be more diffi
cult for drivers to sustain a safe degree of
concentration levels and reaction speed when they are not actively engaged
in driving.6 Human- machine interactions may also become more extensive
and span increasingly complex domains as technologies are developed to
enhance human skills. In the case of robot- assisted surgeries, for example,
physicians may not have enough incentive to obtain suffi
cient training or
to be suffi
ciently prepared for back-up options if the machine were
to mal-
function.
In many of these situations, it may be impractical or too costly for pro-
ducers to monitor individual users and to intervene. Therefore, it would be
important to maintain consumer liability to the extent that users of AI tech-
nologies have suffi
cient incentives to take precautions and invest in training,
thus internalizing potential harm to others. When negative externalities are
suffi
ciently high, regulators may fi nd it necessary to mandate such invest-
ment. For example, a special driver’s license may be required to operate
a self- driving car. Similarly, doctors may be required to take a minimum
number of training sessions with the robotic system before being allowed
to perform certain types of procedures on patients.7
Consumer liability may incentivize users themselves to innovate in ways
that help them take more eff ective precautions (Von Hippel 2005). For ex-
ample, hospitals may redesign the operating room process or reorganize phy-
sicians’ training and work schedules. Furthermore, consumer liability may
also incentivize producer innovation because users would demand safer and
easier- to-use design features (Hay and Spier 2005), and mandatory train-
ing would favor “easier to teach” designs in order to reduce adoption costs.
20.4.2 Federal
Regulation
Another key issue is whether Congress should pass federal regulations on
the safety of AI and robotic technologies that preempts state laws and how
such regulation would aff ect innovation. This would involve the creation of
6. The National Transportation Safety Board determined that the 2016 fatal Tesla crash was partly due to the driver’s inattention and overreliance on vehicle automation despite manufacturer safety warnings.
7. Some expert robotic surgeons and many surgical societies have voiced the need for basic, standardized training and certifi cation in robotic surgery skills (O’Reilly 2014).
500 Alberto Galasso and Hong Luo
a centralized regulatory system similar to the FDA for drugs and high- risk
medical devices: federal regulatory bodies would specify the safety stan-
dards, and approved products would be exempted from state liability claims
under certain conditions.8 For autonomous vehicles, the House passed a