ISSN 1055- 1425
April 2009
This work was performed as part of the California PATH Program of the
University of California, in cooperation with the State of California Business,
Transportation, and Housing Agency, Department of Transportation, and the
United States Department of Transportation, Federal Highway Administration
and conducted by the RAND Corporation.
The contents of this report reflect the views of the authors who are responsible
for the facts and the accuracy of the data presented herein. The contents do not
necessarily reflect the official views or policies of the State of California. This
report does not constitute a standard, specification, or regulation.
Final Report for 6122
CALIFORNIA PATH PROGRAM
INSTITUTE OF TRANSPORTATION STUDIES
UNIVERSITY OF CALIFORNIA, BERKELEY
Liability and Regulation of Autonomous
Vehicle Technologies
UCB- ITS- PRR- 2009- 28
California PATH Research Report
Nidhi Kalra, James Anderson, Martin Wachs
RAND Corporation
CALIFORNIA PARTNERS FOR ADVANCED TRANSIT AND HIGHWAYS
ii
iii
PREFACE
THE RAND ENVIRONMENT, ENERGY, AND ECONOMIC DEVELOPMENT PROGRAM AND
THE RAND INSTITUTE FOR CIVIL JUSTICE
The RAND Corporation is a nonprofit research organization providing objective analysis and effective
solutions that address the challenges facing the public and private sectors around the world. This research was
conducted as a joint effort of the Environment, Energy, and Economic Development Program ( EEED) within RAND
Infrastructure, Safety, and Environment ( ISE) and of the RAND Institute for Civil Justice ( ICJ).
The mission of ISE is to improve the development, operation, use, and protection of society’s essential
physical assets and natural resources and to enhance the related social assets of safety and security of individuals in
transit and in their workplaces and communities. The EEED research portfolio addresses environmental quality and
regulation, energy resources and systems, water resources and systems, climate, natural hazards and disasters, and
economic development— both domestically and internationally. EEED research is conducted for government,
foundations, and the private sector.
Information about the Environment, Energy, and Economic Development Program is also available online
( http:// www. rand. org/ ise/ environ). Inquiries about EEED projects should be sent to the following address:
Keith Crane, Director
Environment, Energy, and Economic Development Program, ISE
RAND Corporation
1200 South Hayes Street
Arlington, VA 22202- 5050
703- 413- 1100, x5520
Keith_ Crane@ rand. org
The mission of ICJ is to improve private and public decisionmaking on civil legal issues by supplying
policymakers and the public with the results of objective, empirically based, analytic research. ICJ facilitates change
in the civil justice system by analyzing trends and outcomes, identifying and evaluating policy options, and bringing
together representatives of different interests to debate alternative solutions to policy problems. ICJ builds on a long
tradition of RAND research characterized by an interdisciplinary, empirical approach to public policy issues and
rigorous standards of quality, objectivity, and independence.
ICJ research is supported by pooled grants from corporations, trade and professional associations, and
individuals; by government grants and contracts; and by private foundations. ICJ disseminates its work widely to the
legal, business, and research communities, and to the general public. In accordance with RAND policy, all ICJ
research products are subject to peer review before publication. ICJ publications do not necessarily reflect the
opinions or policies of the research sponsors or of the ICJ Board of Overseers.
Information about the RAND Institute for Civil Justice is available online ( http:// www. rand. org/ icj).
Inquiries about ICJ research projects should be sent to the following address:
James Dertouzos, Acting Director
RAND Institute for Civil Justice
1776 Main Street
Santa Monica, CA 90407- 2138
310- 393- 0411 x6786; fax: 310- 451- 6979
Robert_ Reville@ rand. org
Questions or comments about this report should be sent to the project leader, Nidhi Kalra
( Nidhi_ Kalra@ rand. org).
iv
v
TABLE OF CONTENTS
Preface........................................................................................................................ .......................................... iii
Table of Contents ............................................................................................................................... ................... v
Abstract ............................................................................................................................... ................................. vii
Executive Summary ............................................................................................................................... .............. ix
Abbreviations.................................................................................................................. ................................... xv
1. Introduction ............................................................................................................................... ........................ 1
2. Autonomous Vehicle Technologies .................................................................................................................. 3
2.1 A Framework for Understanding Autonomous Vehicle Technologies ......................................... 6
2.1.1 Sensing ............................................................................................................................... ...... 6
2.1.2 Planning ............................................................................................................................... .... 7
2.1.3 Acting......................................................................................................................... .............. 8
2.1.4 Integrating Sensing, Planning, and Acting ............................................................................. 8
2.2 The Workings of Autonomous Vehicle Technologies ................................................................... 9
3. Prior Work ............................................................................................................................... ........................ 13
4. The Current Liability Regime and its Application to Autonomous Vehicle Technologies ........................ 17
4.1 Liability for Drivers and Insurers................................................................................................... 17
4.1.1 Theories of Driver Liability................................................................................................... 17
4.1.2 Autonomous Vehicle Technologies and Liability of Drivers.............................................. 19
4.2 Liability of Manufacturers.............................................................................................................. 22
4.2.1 Theories of Liability............................................................................................................... 22
4.2.2 Types of Defectiveness .......................................................................................................... 28
4.3 Effect of Regulation and PreEmption ............................................................................................ 32
4.4 Conclusions and Recurring Themes .............................................................................................. 34
5. Standards and Regulations and Their Application to Autonomous Vehicle Technologies ........................ 37
5.1 Overview of Regulations for Automobiles.................................................................................... 37
5.2 Air bag Regulation .......................................................................................................................... 38
5.3 Current Standards and Regulations for Autonomous Vehicle Technologies.............................. 39
5.4 Future Implications for Standards and Regulations for Autonomous Vehicle Technologies .... 41
6. Summary, Policy Suggestions, and Future Work .......................................................................................... 45
6.1 Summary........................................................................................................................ ................. 45
6.2 Policy Suggestions .......................................................................................................................... 46
6.3 Future Research in Liability and Regulation of Autonomous Vehicle Technologies ................ 47
References..................................................................................................................... ....................................... 51
vi
vii
ABSTRACT
Autonomous vehicle technologies and advanced driver- assistance systems have the potential to
significantly improve transportation safety and efficiency, and, collectively, they may offer tremendous social,
economic, and environmental benefits. As these technologies increasingly perform driving functions, they also
create a shift in responsibility for driving from the driver to the vehicle itself. This motivates a new look at liability
and regulatory regimes because of the increasing uncertainty about what should happen when the inevitable crash
occurs and the implications for the adoption of these technologies. This research is an initial step toward
understanding these issues and creating an integrated collection of policies to address them.
In this work, we first evaluate how the existing liability regime would likely assign responsibility in crashes
involving autonomous vehicle technologies. We identify the controlling legal principles for crashes involving these
technologies and examine the implications for their further development and adoption. We anticipate that consumer
education will play an important role in reducing consumer overreliance on nascent autonomous vehicle
technologies and minimizing liability risk. We also discuss the possibility that the existing liability regime will slow
the adoption of these socially desirable technologies because they are likely to increase liability for manufacturers
while reducing liability for drivers. Finally, we discuss the possibility of federal preemption of state tort suits if the
U. S. Department of Transportation ( US DOT) promulgates regulations and some of the implications of eliminating
state tort liability.
Second, we review the existing literature on the regulatory environment for autonomous vehicle
technologies. To date, there are no government regulations for these technologies, but work is being done to develop
initial industry standards. It will be particularly important for autonomous vehicle technology standards— unlike
standards for most other automotive technologies— to specify precise environmental conditions under which
compliance must be met and tested, include performance standards and tests under a wide range of environmental
conditions, and take into account how diverse populations of drivers might use the technologies.
Keywords: autonomous vehicle technologies, advanced driver- assistance systems, liability, regulation, policy.
vi ii
ix
EXECUTIVE SUMMARY
Autonomous vehicle technologies and advanced driver- assistance systems have the potential to enormously
benefit humankind. Yet, stakeholders have repeatedly voiced substantial concern about tort liability for damages that
may result from the use of these technologies. Who will be responsible when the inevitable crash occurs, and to
what extent? How should standards and regulations handle these systems? This research is an initial step toward
creating policies to address autonomous vehicle technologies.
We define autonomous vehicle technologies as technologies and developments that enable a vehicle to
assist, make decisions for, and, ultimately, replace a human driver. Such technologies include crash warning
systems, adaptive cruise control, lane- keeping systems, and autonomous parking technology ( often collectively
termed advanced driver- assistance systems, or ADASs), and, ultimately, full- scale driverless cars. Collectively,
these technologies can improve safety by reducing the number of crashes, reducing traffic congestion, increasing
fuel efficiency, reducing the environmental effects of driving, and helping to resolve land- use problems ( such as the
need for parking facilities adjacent to homes and businesses).
As these technologies increasingly perform complex driving functions, they also shift responsibility for
driving from the driver to the vehicle itself. This motivates a new look at liability and regulatory regimes because of
the increasing uncertainty about what should happen when the inevitable crash occurs and the implications for the
adoption of these technologies. In this work, we identify the controlling legal principles for crashes involving
autonomous vehicle technologies, evaluate how the existing liability regime would likely assign responsibility in
crashes involving such technologies, and examine the implications for their adoption. We also review the existing
regulatory environment for autonomous vehicle technologies, examine where standards and regulations might fall
short, and suggest general principles and guidelines for future standard setting and rulemaking.
We find that the existing liability regime does not seem to present unusual liability concerns for owners or
drivers of vehicles equipped with autonomous vehicle technologies. On the contrary, the decrease in the number of
crashes and the associated lower insurance costs that these technologies are expected to bring about will encourage
drivers and automobile- insurance companies to adopt this technology.
In contrast, manufacturers’ products liability is expected to increase, and this may lead to inefficient delays
in the adoption of this technology. Manufacturers may be held responsible under several theories of liability for
systems that aid the driver but leave him or her in total or partial control. Warnings and consumer education will
play a crucial role in managing manufacturer liability for these systems, but liability concerns may significantly slow
the introduction of technologies that are likely to increase that liability, even if they are socially desirable. One
potential approach to this problem is to more fully integrate a cost- benefit analysis into the standard for liability in a
way that accounts for the consideration of the benefits associated with this technology. But it is difficult to specify
the appropriate sets of costs and benefits that should be considered, and further research on this issue would be
helpful.
x
xi
One approach to reducing manufacturer liability would be for policymakers ( probably federal) to issue a
uniform set of regulatory standards that would preempt state tort suits. Under the doctrine of preemption, state tort
law claims that are inconsistent with the objective of a federal regulation could be preempted by the federal
regulation. While such an approach offers the advantage of uniform standards, and probably speedier introduction of
these technologies, it also carries significant disadvantages. These include reliance on a single institution to properly
mandate the appropriate standards and the elimination of the traditional access to the courts to seek redress for a
claimed wrong.
As of today, no regulations exist for autonomous vehicle technologies. Looking forward, we find that there
are unique features of autonomous vehicle technologies that future efforts will need to consider. First,
standardization of technology performance and system interfaces will be particularly important for the many
technologies that interact with and share control with the driver. Drivers are likely to use these technologies in
different vehicles created by different manufacturers, and their safe use depends on the driver understanding how to
use each technology and its capabilities and limitations. Second, autonomous vehicle technologies will be used by a
wide range of drivers and passengers who vary in their expectations of how and when the technologies will work
and in their ability to understand each system’s warnings and directions. Therefore, standards must be developed
that take into account diverse populations. Third, autonomous vehicle technologies sense and make judgments about
the vehicle’s external environment, which cannot be controlled and can vary tremendously. Therefore, performance
standards for these technologies must specify the environmental conditions under which the tests must occur and,
ideally, will include testing under a wide range of environmental conditions.
While this report is just a preliminary effort to survey the likely liability and regulatory implications of this
technology, we make a few suggestions that policymakers and automakers may wish to consider. First, it will be
important to ensure that consumer expectations do not exceed the limits of the available technology. While we
anticipate that fully autonomous operation of cars will be possible at some point, the first steps will retain the human
driver as an important part of the control of the car. Drivers will be tempted to rely on the technologies in dangerous
ways. Avoiding false expectations will be important to ensure safety. Manufacturers will have strong incentives to
design systems to prevent this overreliance to reduce liability under legal theories that consider consumer
expectations.
Second, it is important that the operation of these technologies be standardized among manufacturers so
that the technologies function in materially the same way regardless of car manufacturer. This will reduce the risk of
crashes stemming from consumer confusion. Furthermore, significant research has been done on such technologies
as lane departure warning, driver- fatigue monitoring, and collision warning, and the benefits are significant. We
recommend that the US DOT accelerate rulemaking for these mature technologies.
Third, Congress could consider creating a comprehensive regulatory regime to govern the use of these
technologies. If it does so, it should also consider explicitly preempting inconsistent state- court tort remedies. This
may minimize the number of inconsistent legal regimes that manufacturers face and simplify and speed the
introduction of these technologies. While federal preemption has important disadvantages, it might speed the
xii
x ii i
development and utilization of these technologies and should be considered, if accompanied by a comprehensive
federal regulatory regime.
xv
ABBREVIATIONS
3D three- dimensional
ACC adaptive cruise control
ADAS advanced driver- assistance system
CAMP Crash Avoidance Metrics Partnership
CWS collision warning system
DARPA Defense Advanced Research Projects Agency
EEED Environment, Energy, and Economic Development
Program
EPA U. S. Environmental Protection Agency
FDA Food and Drug Administration
FMVSS Federal Motor Vehicle Safety Standard
GNSS global navigation satellite system
GPS Global Positioning System
ICJ RAND Institute for Civil Justice
IHRA- ITS International Harmonized Research Activities working
group on intelligent transport systems
INS inertial navigation system
ISE RAND Infrastructure, Safety, and Environment
ISO International Organization for Standardization
ITS intelligent transport system
IVHS intelligent vehicle highway system
LCDAS lane- change decision- aid system
m Meter
mph miles per hour
NCAP New Car Assessment Program
NGO nongovernmental organization
NHTSA National Highway Traffic Safety Administration
NTSB National Transportation Safety Board
SAE Society of Automobile Engineers
US DOT U. S. Department of Transportation
xv i
1
1. INTRODUCTION
Autonomous vehicle technologies refers to technologies and developments that enable a vehicle to assist, make
decisions for, and, ultimately, replace a human driver. Such technologies include crash warning systems, adaptive
cruise control ( ACC), lane- keeping systems, autonomous parking technology ( often collectively termed advanced
driver- assistance systems, or ADASs), and, ultimately, full- scale driverless cars.
These technologies potentially offer tremendous social, economic, and environmental benefits. In the
United States, more than 40,000 people are killed each year in crashes, and approximately 2.5 million are injured
( NHTSA, 2008c). The vast majority of these crashes are the result of human error ( Choi et al., 2008; FMCSA,
2006), and, by greatly reducing the opportunity for human error, autonomous vehicle technologies have the potential
to greatly reduce the number of crashes. For example, a report by the Federal Highway Administration finds that up
to 48 percent of the 18 million annual rear- end, run- off- road, and lane- change crashes could be prevented with the
widespread deployment of integrated driver- assistance and warning systems ( ITS, undated [ a]). Similarly, a study of
commercial vehicles found that a bundled system of collision warning, ACC, and advanced braking could prevent
23– 28 percent of rear- end crashes ( Battelle, 2007). Because the daily toll from crashes is so high, even small
improvements could save a significant number of lives.
Autonomous driving can also increase fuel efficiency and reduce the environmental effects of driving. One
study found that, “ because [ ACC] reduces the degree of acceleration relative to manual driving, and because [ it]
would be used more than [ conventional cruise control], deployment of [ ACC] systems will result in increased fuel
efficiency and decreased emissions” ( NHTSA, 1999 [ 2002], pp. 5– 17). Bose and Ioannou ( 2001) demonstrated in
simulation that, if 10 percent of vehicles used ACC, fuel consumption could be reduced by 8.5– 30 percent and
pollution could be reduced from 1.5 to 60 percent, depending on acceleration characteristics ( Bose and Ioannou,
2001). Work is currently being done to use these technologies to improve traffic flow ( Kesting et al., 2005; Zhou
and Peng, 2005).
Eventually, occupants will also benefit from being able to safely pursue other activities while their vehicles
drive themselves. Such full- scale driverless cars can offer mobility and independence to those without it. They may
even be used as valets to perform tasks and errands with no driver and no passengers on board. This valet capacity
could also help resolve land- use problems that result from the need for most businesses and residences to have
parking within a short walk. If a car could be quickly summoned, it could be stored at some distance from where it
will be entered or exited.
In some ways, such technologies are not the first “ autonomous” aids to driving: Air bags and antilock
brakes also intervene automatically and have been around for decades. However, emerging and future autonomous
vehicle technologies are unique in the extent to which they contribute to driving. In traditional vehicles, the human
driver is responsible for negotiating virtually all elements of driving, including other vehicles on the road,
pedestrians, road conditions, and the weather. The driver must analyze the highly complex, highly variable
2
environment and is responsible for making judgments about when, where, and how to drive. Autonomous vehicle
technologies will increasingly perform these challenging functions in addition to, or instead of, the driver. In doing
so, such systems increasingly shift responsibility for driving from the driver to the vehicle itself.
This motivates a new look at liability and regulatory regimes because of the increasing uncertainty about
what should happen when the inevitable crash occurs. Who will be responsible and to what extent? In the United
States, state tort law plays a central role in resolving disputes. Accordingly, tort law doctrine will have enormous
implications for the development and implementation ( or lack thereof) of these technologies ( Hunziker and Jones,
1994). Additionally, the government regulates automobiles substantially. Yet, no regulatory standards exist for
autonomous vehicle technologies, despite the fact that some technologies are already on the market. What role
should regulation play, and how should it interact with the relevant tort law in this context? What first principles can
be established in this emerging field?
Such institutional concerns shape investment in autonomous vehicle technologies and, when unresolved,
can pose barriers to their development and adoption ( GAO, 1997; Cheon, 2002; Deakin, 2004). This research is an
initial step toward creating an integrated collection of policies to understand and address these issues. In this work,
we first evaluate how the existing liability regime would likely assign responsibility in crashes involving
autonomous vehicle technologies. We identify the controlling legal principles for crashes involving these
technologies and examine the implications for their further development and adoption. We conclude that
autonomous vehicle technologies are likely to reduce liability for drivers but increase liability for manufacturers as
perceived responsibility for crashes shifts from drivers to the vehicle itself. This may impede development and use
of these technologies. Second, we examine the existing regulatory environment for autonomous vehicle
technologies, examine where standards and regulations might fall short, and suggest general principles and
guidelines for future standard setting and rulemaking. We then offer tentative conclusions about the effect of this
landscape on the introduction of autonomous vehicle technologies and identify areas of potential further research.
We begin in Section 2 with an introduction to autonomous vehicle technologies and discuss prior work in
Section 3. Section 4 presents analysis of the current liability regime and its application to autonomous vehicle
technologies. In Section 5, we discuss standards and regulations and their application to autonomous vehicle
technologies. We conclude in Section 6 with a summary, policy suggestions, and opportunities for future work in
these areas.
3
2. AUTONOMOUS VEHICLE TECHNOLOGIES
We define autonomous vehicle technologies as those technologies that gather information about the
environment and autonomously make judgments about driving to inform the driver, take partial control of the
vehicle, or drive the car autonomously without any human intervention. 1 We have limited the scope of this research
effort to in- vehicle technologies that are not dependent on new infrastructure or the installation of new technology in
the road. 2
A widely used approach to categorizing these technologies is by the degree to which they intervene in the
driving of the vehicle ( Carsten and Nilsson, 2001). This approach is particularly useful for examining the
implications of liability and regulation as control shifts from the driver to the vehicle. However, while much of the
literature uses discrete categories to describe these systems ( e. g., Ho, 2006; Schwarz, 2005), we feel that they are
more aptly described as points along a spectrum, from complete driver control to complete vehicle control. This
notional spectrum is illustrated in Figure 2.1.
1 Many of the technologies we have dubbed autonomous vehicle technologies are often collectively referred
to as advanced driver assistance systems ( ADASs). We use the term autonomous vehicle technologies instead of
ADAS, however, to precisely refer to technologies that gather information and autonomously make judgments about
driving and that, for the purposes of this study, are likely to present a shift in liability. The term advanced driver
assistance systems is imprecise because ( 1) the classification of a technology as advanced is subjective and changes
as new technologies emerge, and ( 2) there is no consensus about what is and what is not an ADAS. Additionally, the
term ADAS does not map well to the technologies we examine. For example, ADAS typically includes technologies
that only present information to the driver without making judgments, such as navigation aids and night- vision aids,
and excludes such technologies as fully autonomous vehicles.
2 We focus our analysis on technologies in which the driving “ intelligence” resides within the vehicle rather
than in the surrounding road infrastructure ( i. e., intelligent highways) for two reasons. First, as we discuss in our
review of prior work in Section 3, infrastructure- based systems were the focus of much of the earlier research and
development efforts, and, consequently, many others have already discussed the liability and regulatory regimes for
these technologies ( e. g., Ayers, 1994; Cheon, 2002; Khasnabis, Ellis, and Baig, 1997; Roberts et al., 1993; Syverud,
1992). Second, and perhaps more important, in- vehicle systems are reaching drivers much more rapidly and widely
than infrastructure- intensive systems, which must overcome a wide range of institutional challenges ( see Deakin,
2004; Cheon, 2002; Alkadri, Benouar, and Tsao, 1998). In 1997, for example, the U. S. Department of
Transportation ( US DOT) discussed a shift in the National Automated Highway System Consortium’s focus on
driver- assistance systems because of a widespread lack of support among administration and other stakeholders for
the high- cost, long- term planning horizon and liability concerns associated with the automated highway system
( GAO, 1997). This program was terminated early in 1998. Given the lack of political support for expensive,
wholesale changes in the U. S. transportation infrastructure, the liability and regulation of the many in- vehicle
systems appearing on the market more immediately affect consumers and manufacturers. This is not to suggest that
infrastructure is unnecessary— the Global Positioning System ( GPS), for example, is a critical navigation component
for many autonomous vehicle technologies. Moreover, much research is under way in both the public and private
sectors to develop infrastructure that assists and enables in- vehicle autonomous vehicle technologies, such as
communication networks for vehicle- to- vehicle communication ( Shulman and Deering, 2005). Similarly,
researchers have examined pathways to pilot and then broadly adopt these systems ( e. g. Shladover et al., 2001). A
new look at liability and regulation of these systems will be warranted as these support systems are deployed and as
new public- private partnerships emerge to manage them.
4
Near one end of the spectrum, warning and assistance systems alert the driver to dangerous conditions,
such as unintended lane departures and impending forward, rear- end, and lateral collisions, and guide the driver in
maneuvers, such as changing lanes and parking. Simple forward collision- avoidance systems, for example, monitor
the speed differential with and the distance from the preceding vehicle and alert the driver with an audio cue when
there is a risk of a collision. Other technologies aim to protect those outside the vehicle. Pedestrian- detection
systems, for example, identify and alert drivers to pedestrians and cyclists, particularly if they are in low- visibility
areas toward the rear of the vehicle or in the driver’s blind spot ( Chan, 2006). Such technologies affect the driving
process indirectly by influencing the driver rather than by acting on the vehicle. In this way, the driver must interpret
the suggestions and warnings, take into account other factors that may affect the validity of the information ( e. g., the
weather, other traffic, road conditions), and, ultimately, drive the vehicle.
Intervention systems take the next step toward vehicle autonomy. They gather information, process it, and
take partial control of the driving task in some way. ACC is perhaps the best known of these technologies.
Traditional cruise control fixes the velocity of the vehicle without regard to the external environment, while ACC
tailors cruise control to the flow of traffic: It reduces speed when preceding vehicles are moving slower than the
desired speed and then accelerates again when becomes safe to do so. In this way, ACC is able to reduce forward
collisions, particularly when the driver may be distracted and where the flow of traffic is uneven. Lane- keeping
systems use vision or laser technologies to track lane markings and keep the vehicle in its current lane. Although
these technologies actively control the vehicle in some ways, they must be activated by the driver, with the idea that
the driver is responsible for judging when it is appropriate to use them ( e. g., when traffic and weather conditions
Figure 2.1. Notional Spectrum of Control. It is common to describe and categorize autonomous vehicle
technologies according to the degree to which they intervene in the driving of the vehicle. At one end of this
spectrum, conventional vehicles are completely in the driver’s control, while, at the other end, driverless cars do not
require a human behind the wheel. Note that the distances between systems in this figure are for visual clarity only;
they do not reflect “ distances” in terms of technical complexity.
Conventional
vehicles
Driverless cars
Driver- warning
systems ( e. g., forward
collision warning)
Partial control systems
( e. g., ACC and precrash
safety)
Short- term driving
systems ( e. g., ACC and
lane keeping combined)
Autonomous vehicles
with driver- in- the-loop
5
allow). However, some of these technologies disengage automatically. Many ACC systems, for example,
automatically deactivate at low speeds when cruise control may not be appropriate: They warn the driver that
intervention is required and then release control of the braking and acceleration with the expectation that the driver
will resume control over those functions.
A few technologies exert partial control of the vehicle without being engaged by the driver. To date, such
systems are limited to those that intervene when a crash is imminent and no action on the part of the driver can
prevent it. For example, precrash safety systems tighten seat belts and apply the brakes prior to impact in an effort to
reduce the effects of the crash ( e. g., DENSO Corporation, undated). Parallels can be drawn between these
technologies and air bags: Both activate without the driver’s influence but only when a crash is imminent and with
the aim of reducing injury and damage. However, a distinction can be made between air bags and these
technologies: Air bags are not controlling the central functions of the car and are triggered by the impact itself.
Next, we can foresee a host of systems that take over driving for short periods of time or in certain
circumstances at the discretion of the driver. For example, combining ACC and lane keeping creates a system that is
capable of driving itself in highway- like environments: ACC maintains speed, and lane keeping maintains steering. 3
Similarly, other technologies may autonomously park the vehicle or change lanes on the highway. As before, the
driver will ( at least in the near future) have to activate these systems at his or her discretion.
Finally, at the far end of the spectrum are fully autonomous vehicles that are capable of driving from origin
to destination entirely on their own. Such vehicles are not yet available, but significant leaps forward have been
made in recent years. The Defense Advanced Research Projects Agency ( DARPA), for example, recently held a
series of high- stakes, high- profile competitions among industry and university research groups to develop fully
autonomous vehicles, first for off- road and trail driving in the Grand Challenge and then, in 2007, for urban
environments in the Urban Challenge. The vehicles completed a series of driving missions completely autonomously
with no human intervention while obeying all traffic rules and operating in the presence of other autonomous and
human- driven vehicles ( Urmson et al., 2007).
We can envision several roles for fully autonomous passenger vehicles. First, they could enable would- be
drivers to perform other tasks, turning morning commutes, for example, into opportunities to catch up on sleep,
enjoy breakfast, or prepare for the day’s work. In such scenarios, the driver is in the vehicle but is not involved in
the driving process in any way. Second, they would provide mobility and independence to those without it,
including the elderly, the physically challenged, and even children. The vehicle, for example, could chauffeur the
elderly to medical appointments or children to school and soccer practice. Here, passengers are in the vehicle but
they are not intended to participate in driving. Finally, vehicles would be able to act as valets with no driver and no
passengers on board. The vehicle could pick up dry cleaning and groceries and return books to the library or park
itself after dropping the owner off at work. These scenarios demonstrate different levels of human involvement in
3 In 2006, Honda launched its Accord ADAS for sale to the public in Europe with just such capabilities
( Miller, 2006).
6
personal transportation, even when the vehicle is fully autonomous; thus, they may also present different liability
concerns. As we discuss later, crashes ( and liability) may result if consumers are confused about the capabilities of
technology. This has heightened concern for technologies that only partially take control of the automobile and still
leave the human ( and, therefore, error- prone) driver in the control loop.
2.1 A FRAMEWORK FOR UNDERSTANDING AUTONOMOUS VEHICLE TECHNOLOGIES
In traditional vehicles without autonomous vehicle technologies, the human driver negotiates virtually all
of the interaction between the vehicle, the driving objective, and the external environment. The driver determines the
long- term navigation task ( e. g., arriving at the grocery store in 20 minutes), continuously evaluates the environment
surrounding the vehicle ( including other vehicles, lane markings, signs, pedestrians, weather, and road conditions),
determines the desired vehicle behavior ( e. g., to change lanes), decides how to achieve that trajectory ( e. g., a smooth
arc or a sharp swerve), and exerts control over the vehicle by manipulating lateral movement and speed. The driver
is also responsible for providing information to others in the environment— e. g., waving another driver through an
intersection, honking at a pedestrian who appears to be walking into the line of traffic, or using signals to indicate a
turn. The vehicle’s only “ autonomous” response to the environment occurs when the environment has an immediate
effect on the vehicle’s performance. For example, the vehicle’s air bags automatically deploy only when its crash
sensors measure the actual force of an impact but not in the split second before, when the impact is imminent and
perhaps anticipated by the driver but not yet experienced by the vehicle. Similarly, antilock braking systems engage
when the wheel speed sensors determine a disparity in rotational speeds between the four wheels that is indicative of
a locking wheel but do not sense or respond directly to the driving conditions that may have caused it ( e. g., icy
roads).
In contrast to technologies in conventional vehicles, autonomous vehicle technologies perform many of
these complex functions and respond to the environment directly; this enables them to play ever- larger roles in
driving. The robotics research literature roughly categorizes these functions into three categories: sensing, planning,
and acting. This framework is an oversimplification and an overgeneralization of the processes unique to each
technology and each implementation, but the terms can help us understand how these systems operate. Sensing,
planning, and acting are functions that humans perform naturally and seamlessly, usually without conscious thought.
For autonomous systems, however, each can be an extremely complex and difficult engineering challenge. We begin
by describing these different functions and how they are performed, and then examine different autonomous vehicle
technologies.
2.1.1 Sensing
Sensing consists of gathering information about the external environment and the internal system in order
to build a model, called a world model, that represents and describes the vehicle, its surroundings, and the
relationships between them. Depending on the system function, this model may include the external environment,
such as road conditions, weather, and traffic; vehicle- performance characteristics, such as velocity, heading, and tire
7
pressure; and even the behavior of vehicle occupants, such as the driver’s eye movements, seat- belt use, and
passenger weight distribution. In a fully autonomous vehicle, all of this information may need to be incorporated
simultaneously into a complete world model.
There are a number of challenges associated with sensing, particularly for systems that perform complex or
multiple functions. First, individual sensors are limited in what they can detect or measure and depend on favorable
environmental conditions. 4 Consequently, systems may need to have many different sensors to gather all the
required information and to provide redundancy and increased reliability. For example, Tartan Racing, a team in the
DARPA Urban Challenge, employed eight different types of sensors on its vehicle and multiple sensors of each type
( Urmson et al., 2007). These sensors generate a tremendous amount of data per second. Second, the vehicle must be
able to process the data fast enough to avoid a backlog of information. Third, some of the data will be good data
( e. g., color information from cameras in the day), and some not ( e. g., color information from cameras at night), and
the system must be able to recognize the difference. Fourth, data from different sensors or gathered at different times
may conflict; algorithms must reconcile contradictions in the data and, in the end, create a complete model that is
accurate enough to enable the vehicle to drive safely and efficiently.
2.1.2 Planning
In planning, the system applies an algorithm to analyze the world model and create a set of actions that
bring the system closer to its objective. In complex systems, such as fully autonomous vehicles, planning produces a
sequence of high- level behaviors similar to those a human would take ( e. g., change lanes, then take the next exit and
turn right at the end of the ramp) and trajectories to achieve those behaviors ( e. g., a sequence of wheel curvatures
and speeds). In simpler systems, however, this step may be omitted if features in the world model can be mapped a
priori onto desired actions ( e. g., in a precrash braking system, the brakes may engage whenever the distance to the
preceding vehicle is less than 2 meters [ m] and the vehicle speed is more than 10 miles per hour [ mph]).
4 Vision, radar, and laser are some of the most common sensor types for measuring the external environment.
Vision- based systems use cameras to detect color and shape but depend on favorable lighting conditions. Used in
stereo pairs, cameras can also determine distances to objects but with less accuracy than most laser and radar
systems. Laser and radar emit light and radio signals, respectively, which reflect off of objects in the environment
and return to the sensor. By measuring the time of the signal flight, these sensors can determine the distance to and
velocity of objects and construct three- dimensional ( 3D) representations of the environment. Laser systems measure
distances to all types of solid objects, but they are best at short range and do not operate on water or in rain. Radar
systems, on the other hand, operate at long ranges but are limited to detecting metallic objects; this is ideal for
detecting vehicles but not other traffic.
GPS and inertial navigation systems ( INSs) are the most common sensors for detecting the vehicle’s own
position, velocity, and orientation. GPS is a global navigation satellite system ( GNSS) that provides accurate
absolute positioning, orientation, and velocity by measuring signals from satellites orbiting the earth. As such, GPS
requires a clear view of the sky and does not perform well in urban environments or under tree cover. INSs work in
all environments: They use gyroscopes and accelerometers to accurately track the vehicle’s position, orientation, and
velocity relative to a starting reference point and are critical when external sources of position ( e. g., GPS) are
unavailable. INSs, however, can accrue significant error over time.
8
One of the challenges in planning is that the system needs to select actions that will occur in the future even
when it has accurate information only about the past and the present. Therefore, it needs to extrapolate the future
state of the environment, which may contain moving elements that often do not behave as expected. For example, a
vehicle’s plan may involve it driving at a particular speed and in a particular direction, but this may be based on the
assumption that the preceding vehicle will maintain its current speed. If the preceding vehicle suddenly decelerates,
the system will have to very rapidly prepare a new plan of action. This is a second challenge in planning: The
system needs to compute plans fast enough to always be responsive to the driving task, which can require several
planners to be running simultaneously. 5
2.1.3 Acting
Lastly, acting is the actual execution of plans by manipulating the steering, throttle, brakes, and other
vehicle components. Vehicles today are increasingly being developed with drive- by- wire technology, which
replaces the traditional mechanical and hydraulic control systems with electronic control systems. This allows
autonomous vehicle technologies to use computers to control the vehicle components, which is a far easier and more
reliable approach than manipulating the components with physical force. Although drive- by- wire tremendously
simplifies acting on the vehicle components, there are still challenges associated with producing the desired driving
effect. For example, on an icy road, turning the wheel column rapidly may turn the vehicle’s wheels but may not
turn the vehicle itself because of a loss of traction. A tight sensing, planning, and acting cycle makes it possible to
detect and correct unexpected vehicle behaviors and unexpected traffic and environmental events.
2.1.4 Integrating Sensing, Planning, and Acting
In classic robotic control architecture, sensing, planning, and acting are distinct processes that take place in
sequence. Similar approaches may work in simple autonomous vehicle technologies— e. g., those that warn the
driver when specific conditions are met. In systems like fully autonomous vehicles, however, the difficulty and
uncertainty of the driving task requires that these functions be performed continuously and asynchronously and be
tightly integrated in multiple control loops. For example, there may be a very fast sense- act control loop between a
forward- looking laser and the brakes that slows or stops the vehicle whenever a close obstacle is encountered,
regardless of the higher- level transportation goals ( e. g., changing lanes). The data from this same laser may
simultaneously feed into a larger sense- plan- act control loop to perform other maneuvers. Not surprisingly, these
systems can become very complex, and their performance and reliability can be difficult to test and prove.
5 Tartan Racing’s explanation of its approach to planning describes these challenges in detail ( Ferguson et
al., 2008).
9
2.2 THE WORKINGS OF AUTONOMOUS VEHICLE TECHNOLOGIES
Building on this foundation, we now turn to how specific autonomous vehicle technologies work, paying
particular attention to those aspects that may have implications for liability and regulation. Driver- warning systems
are intended to perform very specific functions: monitor immediate areas around the vehicle and alert the driver
when certain risky conditions occur. Collision warning systems ( CWSs), for example, continuously sense the
distance to the vehicle immediately ahead or behind of the host vehicle, typically with a camera or radar, and
determine when that vehicle is within some distance threshold that poses a threat. Lane departure warning systems
sense the lane markings of the host vehicle’s lane with cameras or lasers and alert the driver when the vehicle
appears to be veering from its lane unintentionally. Indications that the lane change is unintentional may include that
the driver has not turned on the indicator or that the driver’s gaze is not directed at the road as determined by an
inward- looking camera that monitors the driver’s eye movement.
From a technological perspective, such systems are relatively simple. The world model required for
operation is small, consisting of the vehicle, its lane, and vehicles immediately ahead or behind. This means that
sensing is limited to one or two sensors that monitor very specific areas of the road and that it is relatively easy to
fuse sensor data. These systems are also simple in terms of their decisionmaking: Instead of planning, it suffices to
evaluate whether the world model meets some predefined conditions that constitute a scenario that warrants a
warning. Further, as warning systems, they also do not act on the vehicle.
Systems that exert some control on the vehicle use technologies and approaches similar to those in warning
systems. ACC senses the distance to the preceding car using the same technology as forward collision warning
systems and uses the world model to compute the appropriate speed adjustment required to meet the vehicle spacing
set by the driver. Drive- by- wire technology makes it possible to compute and execute smooth throttle adjustments.
In this way, ACCs use a sense- act model and do not need to do any planning. A lane- keeping system is to a lane
departure warning systems what ACC is to a forward collision warning system: it utilizes the same types of sensors
and world models to compute desired vehicle headings and to execute smooth changes in heading.
One challenge is to ensure that the driver understands when the system works properly and when it could
fail or has failed. These systems may not operate in certain weather or visibility conditions in which sensors are
unable to operate or in circumstances that may “ confuse” the sensor. For example, lane departure warning systems
may not effectively detect lane departures in areas of active construction, where several lane markings may overlap
to indicate lane shifts or may be missing entirely. In these circumstances, what is the extent of a manufacturer’s
responsibility in alerting the driver when the system may not be operating properly, given that the driver is expected
to be continuously attentive? What level of system integrity is required? 6 Additionally, for some systems, the driver
is expected to intervene when the system cannot control the vehicle completely. For example, if a very rapid stop is
required, ACC may depend on the driver to provide braking beyond its own capabilities. ACC also does not respond
6 We use the term system integrity to describe a system’s ability to recognize that it is performing unreliably
( e. g., because its components have malfunctioned or because it is operating in an unsuitable environment) and to
notify the user that it should not be used.
1 0
to driving hazards, such as debris on the road or potholes— the driver is expected to intervene. Simultaneously,
research suggests that drivers using these conveniences often become complacent and slow to intervene when
necessary; this behavioral adaptation means drivers are less responsive and responsible than if they were fully in
control ( Rudin- Brown and Parker, 2004). Does such evidence suggest that manufacturers may be responsible for
monitoring driver behavior as well as vehicle behavior? Some manufacturers have already taken a step toward
ensuring that the driver assumes responsibility and is attentive, by requiring the driver to periodically depress a
button or by monitoring the driver by sensing eye movements and grip on the steering wheel. As discussed later,
litigation may occur around the issue of driver monitoring and the danger of the driver relying on the technology for
something that it is not designed to accomplish.
The complexity of systems that take full control of the vehicle at the discretion of the driver depends on
what functions these systems perform. The underlying “ autopilot” system resulting from the combination of ACC
and lane keeping may be no more complex than the sum of its individual parts: It may still utilize a sense- act
paradigm that uses relatively few and simple sensors to build a basic world model that is governed by
straightforward behavioral rules. However, the combination of ACC and lane keeping could cause consumers to
overrely on the system in a way that they would not if they were using each system independently. Then, new
additional technology may need to be introduced to counteract this reliance ( e. g., by monitoring the driver
carefully). In this way, the system may become be much more complex than the sum of its parts.
Such systems as autonomous parking and lane changing will be more complex. They will need to look
beyond the narrowly defined vehicle environment that has sufficed for other systems. Automatic parallel- parking
systems, for example, have to sense the stationary vehicles around the parking space and pedestrians who may be on
the sidewalk; monitor the trajectory of cars, bicycles, and other objects coming up from behind and approaching
from ahead; and plan complex maneuvers to fit in the space.
Fully autonomous vehicles, such as the entrants into DARPA’s Grand and Urban Challenges, are
tremendously more complex. As discussed earlier, they must sense the environment broadly, using a host of sensors
that may present conflicting or incomplete pictures of the environment; develop complex current and future world
models; rapidly plan and replan given uncertainties and constantly changing circumstances; and maintain total
control over the vehicle. The challenges of these technologies, however, go beyond system and algorithmic
complexity. While it is possible to design autonomous vehicles that obey all traffic rules, they, like human drivers,
must share the road with other drivers who do not do so. Additionally, there are many unique human aspects to
driving: Human drivers watch the body language and eye contact of pedestrians to anticipate whether they will
attempt to cut into traffic, drivers signal to each other with eye contact and body language, and local driving
etiquette can preempt actual traffic rules. 7 How can technology be designed to accommodate these customs? What
7 In Pittsburgh, Pennsylvania, for example, the “ Pittsburgh left” is a regional custom in which the first driver
stopped at a traffic light will sometimes wait for a moment after the light turns green to yield to the first driver in the
oncoming traffic who wishes to make a left turn. In most other places, of course, drivers wishing to make a left turn
across traffic yield to the oncoming traffic.
1 1
happens when others expect customary behaviors from drivers that are not implemented in driverless cars?
Furthermore, in many driving situations, the correct procedures are often unclear. If a vehicle breaks down in one
lane of a two- way street, drivers coming in either direction will usually take turns to pass and keep the traffic
flowing. Negotiating such situations involves taking some risks and making some assumptions. Can autonomous
vehicles be designed to behave safely and appropriately in such situations? Experiences from the Urban Challenge
suggest that it is possible: During the competition, some vehicles were able to take appropriate mitigating actions
when others were not operating correctly. But, what risks and assumptions can autonomous vehicles make, and will
we and should we allow it? Who will be responsible under these circumstances?
1 2
1 3
3. PRIOR WORK
Before beginning our own analysis, we briefly discuss prior work that addresses the implications of liability
and regulation on the adoption of autonomous vehicle technologies and closely related systems.
Intelligent vehicle highway systems ( IVHSs) received significant attention in the scientific research
community in the 1990s, and a complementary body of work addresses the liability issues that arise from their
deployment. Much of the technical research emphasized infrastructure- intensive intelligent and automated
highways. Given the government’s major role in implementing such systems, the liability research focuses on
government liability and sovereign immunity, a doctrine that protects federal, state, and local governments from
being held liable.
Roberts et al. ( 1993) thoroughly cataloged the status at the time in the 50 states of the doctrine of sovereign
immunity, a doctrine that sometimes immunizes governmental agents from tort liability, and list ( but do not analyze)
several options that could be used to create an environment that encourages IVHS development without
compromising safety. Khasnabis, Ellis, and Baig ( 1997) provide an initial look at how sovereign immunity applies
to different aspects of government involvement in automated highway systems ( e. g., planning, design, development)
and pose the question, “ How can the government balance the interest of citizens at large against protecting its
agencies, employees, and contractors from lawsuits?” Balancing safety and innovation is still a key issue for the
introduction of autonomous vehicle technologies and motivates our own work; however, this earlier body of work
does not address the host of new technologies that focus on driver assistance and autonomy within individual
vehicles. This move to decentralized networks of autonomous vehicles makes issues of sovereign immunity less
relevant because the requirement for government action is greatly lessened.
Some of the early research also addresses the then- emerging driver- assistance systems, particularly
collision warning and ACC. In a series of three articles, Bagby and Gittings discuss general tools for highway
liability risk management ( 1999a); explain principles of tort reform, products liability reform, and privacy law
( 1999b); and apply these principles to advanced highway systems and electronic toll collection ( 2000). The third
article additionally identifies key litigation risks for collision warning and intervention systems ( e. g., sensor
malfunctions that cause abrupt and unnecessary stopping; system failure to warn drivers of potential collisions;
driver overreliance on systems) and for ACC ( e. g., outright malfunction, such as a failure to disengage; driver’s
inability to control the system or intervene when necessary).
Ayers ( 1994) surveyed a range of emerging autonomous vehicle technologies and automated highways,
evaluated the likelihood of a shift in liability occurring, discussed the appropriateness of government intervention,
and highlighted the most- promising interventions for different technologies. Ayers found that collision warning and
collision- avoidance systems “ are likely to generate a host of negligence suits against auto manufacturers” and that
liability disclaimers and federal regulations may be the most effective methods of dealing with the liability concerns
( p. 21). The report was written before many of these technologies appeared on the market, and Ayers further
1 4
speculated that “ the liability for almost all accidents in cars equipped with collision- avoidance systems would
conceivably fall on the manufacturer” ( p. 22), which could “ delay or even prevent the deployment of collision
warning systems that are cost- effective in terms of accident reduction” ( p. 25).
Syverud ( 1992) examines the legal cases stemming from the introduction of air bags, antilock brakes,
cruise control, and cellular telephones to provide some general lessons for the liability concerns for autonomous
vehicle technologies. In another report, Syverud ( 1993) examines the legal barriers to a wide range of IVHSs and
finds that liability poses a significant barrier particularly to autonomous vehicle technologies that take control of the
vehicle. In this work, Syverud’s interviews with manufacturers reveal that liability concerns had already adversely
affected research and development in these technologies in several companies. One interviewee is quoted as saying
that “ IVHS will essentially remain ‘ information technology and a few pie- in- the sky pork barrel control technology
demonstrations, at least in this country, until you lawyers do something about products liability law’” ( 1993, p. 25).
Syverud sums up that, “[ a] bsent a substantial shift of responsibility for automobile accidents or significant
legislative reform, tort liability is likely to deter significant private investment in [ advanced vehicle- control systems]
in the United States” ( 1993, p. 51).
Collectively, this body of work frames the general concerns about liability and autonomous vehicle
technologies, identifies specific types of litigation that could be expected in the then- emerging technologies, and
suggests initial strategies for addressing the risks. This work provides a foundation for the analysis in this report, but
it is now more than a decade old. Written at a time when autonomous vehicle technologies were just beginning to be
developed, this work is necessarily speculative and does not consider recent advances in or actual implementations
of autonomous vehicle technologies or recent developments in tort law.
Some recent work has been done on the legal implications of ADASs, though most of this work is from
Europe and relates specifically to European law. The RESPONSE 1, RESPONSE 2, and RESPONSE 3 projects are
sponsored by the European Commission and the European auto industry to develop a code of practice for
accelerating the market introduction of ADASs. Part of this work is to examine the liability risk of these
technologies, but the work to date is very general and does not yet include the legal analysis that is necessary
( Schwarz, 2005; RESPONSE 2, 2004). 8 Van Wees and colleagues have published a number of articles ( van der
Heijden and van Wees, 2001; van Wees, 2000, 2004; van Wees and Brookhuis, 2005) mapping out the likely
liability consequences of some categories of ADAS- related crashes under European law. Like others, van Wees and
Brookhuis note, “ One of the most prominent questions in this context is to what extent the use of ADAS may shift
liability for road crashes from drivers to car manufacturers and system developers and whether current liability
regimes are suited to accommodate the introduction of ADAS” ( p. 358). Their findings, based on the European
Products liability Directive, include that manufacturers must take care to educate consumers about technology
8 For example, Schwarz ( 2005, p. 3) observes that “[ t] here is a possibility that the information provided by
the system may be incorrect or inaccurate. If this is the case, manufacturer or distributor liability should also be
taken into consideration.”
1 5
limitations, to consider the behavior of a range of users when defining “ reasonable use” of the technology, and to
test technologies in a wide range of circumstances. While these guidelines will be generally relevant to
manufacturers in the United States, the legal analysis cannot be directly transferred from Europe. Van der Heijden
and van Wees ( 2001, p. 321) note that, “ Due to [ several] differences [ in] the legal situation in the field of products
liability law in Europe and the USA, although in terms of the underlying principle [ they are] the same, in practice
[ they are] not comparable.” When comparing the European and U. S. regimes, van Wees ( 2000, p. 7) found that
“ products liability in the US will indeed be a greater threat for the deployment of electronic driver support systems”
than in Europe and attributed this to broad legal differences, such as high damage awards, the use of technical expert
testimony, and the system of jury trials in the United States, rather than fundamental differences in products liability
laws.
Recent research and development efforts undertaken in the United States do recognize the need for
resolving liability, regulation, and other nontechnical concerns. For example, the US DOT’s IntelliDrive ℠ program
aims to enable vehicles to identify roadway threats and communicate these threats to other vehicles over a wireless
network; its work plan includes the development of wireless- communication standards and studies of liability and
regulation ( see ITS, undated[ b]). Similarly, participants at the 2004 workshop of the US DOT’s Cooperative
Intersection Collision Avoidance Systems program have expressed concerns about the distribution of liability among
the many stakeholders ( see ITS, 2005).
Nevertheless, recent work on these issues seems to only broadly and theoretically survey the liability
concerns associated with of autonomous vehicles and fails to offer a concrete analysis. Cheon’s ( 2002) examination
of the institutional barriers to automated highway systems highlights general liability issues that also apply to in-vehicle
autonomous vehicle technologies. The observations include that liability concerns may lead to highly
conservative regulation and practice, that tort reform may be necessary, and that consumer education and the
development of standards may reduce liability. While this work frames many of the concerns, it does not provide a
specific analysis of the liability regime. Ho ( 2006) explores the policy implications of driver- warning systems and
pays particular attention to standards and regulations, in part as a way to address liability concerns. Like Cheon’s,
however, this work does not include a legal analysis. Nevertheless, Ho ( 2006, p. 100) finds that the benefits of safety
standards outweigh the costs and that “ safety standards must be designed and set as soon as possible in order to
ensure that future warning systems meet a minimum level of safety before they are introduced more widely into the
market.”
In sum, the literature substantiates the importance of rational, well- developed regulatory, safety, and
liability regimes for autonomous vehicle technologies and offers some building blocks to this end. However, it
simultaneously highlights that an analysis of the liability and regulatory implications of autonomous vehicle
technologies has not been conducted.
1 6
1 7
4. THE CURRENT LIABILITY REGIME AND ITS APPLICATION TO AUTONOMOUS
VEHICLE TECHNOLOGIES
In this section, we explain the principles of motor vehicle crash and products liability law and how they are
likely to apply to autonomous vehicle technologies. While a full discussion of the goals of tort law lie well beyond
this work, in general, we view the role of a liability regime as providing appropriate incentives to encourage
efficient precautions against crashes, compensating victims, and promoting corrective justice. 9 This, of course,
includes the efficient development of technology that can reduce crashes.
We do not anticipate liability of drivers ( and their insurers) for auto crashes to delay the introduction of
autonomous vehicle technologies. Indeed, we anticipate that the reduction in crashes caused by human error may
encourage adoption of these technologies. Manufacturers, in contrast, may face increased liability under principles
of products liability law and changed social expectations about the cause of automobile crashes. This may lead to
inefficient delays in the integration of this technology into automobiles.
4.1 LIABILITY FOR DRIVERS AND INSURERS
The liability implications of autonomous vehicle technologies are largely defined by the body of law that
governs motor vehicle crashes and products liability law. The law governing crashes is a mixture of state tort law
and the state’s financial responsibility law that mandates insurance for drivers. As a result, the mandatory- insurance
regime substantially affects litigation that occurs after crashes. We begin by discussing driver liability and then
apply these theories of liability to autonomous vehicle technologies.
4.1.1 Theories of Driver Liability
There are three basic theories of tort liability that affect drivers— traditional negligence, no- fault liability,
and strict liability— and we discuss each in turn. 10
Under traditional negligence principles, a person is civilly liable for harms that the person causes if the
harm is a tort. The wrongdoer ( tortfeasor) must compensate the victim for the harms suffered. 11 The traditional
elements of a negligent tort are the existence of a duty, the breach of that duty, causation, and injury. In the case of
9 By efficient, we generally mean that the social benefits of the technology outweigh the social costs. One
commonly understood ( but not uncontroversial) function of tort law is to encourage the undertaking of safety
precautions to efficiently minimize risk. A sizable literature exists explicating this function of tort law. See, e. g.,
Calabresi ( 1970).
10 Tort is a general term for wrong and is also a term that refers to the general legal field of torts. A negligent
tort is a particular category of tort. Other categories include intentional tort ( the wrongdoer deliberately harms the
victim) and strict- liability torts ( the wrongdoer is liable for harms regardless of whether he or she took reasonable
care to prevent them).
11 In some cases, the tortfeasor must pay punitive damages in addition to those necessary to compensate the
victim ( compensatory damages).
1 8
automobiles, drivers have a duty to take reasonable care in operation. Drivers are liable for injuries that they cause in
violation of this duty of reasonable care.
The central idea of liability for negligence is that a party should be held liable for harms caused by it
unreasonably failing to prevent the risk. For example, suppose that Marty drives a car with defective brakes because
he is too lazy to get the brakes repaired. As a result of his defective brakes, he injures a pedestrian. In this
hypothetical, Marty will probably be found negligent. It was not reasonable for Marty to fail to repair his brakes.
For good and bad, reasonableness is a fairly vague concept. Judge Learned Hand attempted to provide
more guidance by defining negligence with a mathematical formula: B < PL ( United States v. Carroll Towing, 159
F. 2d 173, 1947). In Hand’s formulation, B, the proper burden of care, was defined by the probability of harm, P,
multiplied by the cost of the loss that would result, L. If a measure to prevent the harm cost more than the cost of the
harm multiplied by the probability of the harm, it would not have been efficient for the defendant to have taken and
the defendant should not be considered negligent. If, on the other hand, the accident- reducing measure cost less than
the probability of harm multiplied by the cost of the harm, it would have been efficient to undertake the precaution
and the defendant should be found negligent for failing to do so. In this way, Judge Hand attempted to provide an
economic metric and conceptual rigor to the analysis of the accident- prevention methods— the duty of care—
required. In Marty’s case, let us assume that the cost of the pedestrian’s injury was $ 10,000 and that Marty’s chance
of injuring the pedestrian was 50 percent if he did not repair the brakes: 0.5 × $ 10,000 = $ 5,000. If the cost of
repairing the brakes was less than $ 5,000, Marty should be found liable, using Judge Hand’s calculus. 12
In the day- to- day resolution of automobile crash claims, the operation of the traditional system of liability
for negligence has been influenced by the mandatory- insurance system. Insurance adjusters have adopted informal
rules to effectively allocate fault ( e. g., drivers who rear- end other vehicles are presumed to be at fault). These have
minimized more- general analyses of reasonableness and causation in most automobile crash cases, which are
resolved without formal litigation. So rather than undertaking a generalized analysis of whether a driver is negligent
and therefore liable for a crash— a potentially difficult and open- ended inquiry— an insurance adjuster is likely to
refer to a simpler set of rules to determine who owes what to whom ( Hensler et al., 1991; Ross, 1980, p. 237: “ The
law of negligence was made to lean heavily upon the much simpler traffic law”).
Twelve states currently use an alternative system, called no- fault, for automobile- crash litigation and
insurance. In these states, automobile crash victims are not permitted to sue other drivers in the tort system unless
their injuries reach a certain degree of severity, called a threshold. 13 Instead, victims recover their losses through
12 Applying Judge Hand’s formula in practice raises a host of issues. Which costs and benefits should we
include in the calculus? See Anderson ( 2007, arguing that economic analysis of tort law is indeterminate without a
theory of which variables should be in the cost- benefit calculus).
13 States have adopted either a monetary or verbal threshold. A monetary threshold is a certain dollar amount
that the victim’s injuries have to exceed to recover in tort ($ 5,000 in many states). A verbal thresholds is a
description of a certain degree of seriousness. In Pennsylvania, for example, a plaintiff who elects the limitation on
tort when purchasing automobile insurance must show that his or her injury is “ serious” in order to recover in tort
1 9
their own insurance, which directly compensates them for their losses. Proponents of this system argued that it
would eliminate the difficult determination of who, if anyone, was at fault for a particular crash, ensure that
compensation would be available to victims regardless of whether anyone was legally at fault, and generally reduce
litigation and lawsuits. 14
A rare theory of liability that might also affect operators of autonomous vehicle technologies is strict
liability for abnormally dangerous or ultrahazardous activities. The rationale for this type of liability is that actors
involved in highly unusual activities are more knowledgeable about the risks that such activity entails and should
consequently bear the associated costs regardless of whether they are legally at fault for the crash. 15 This theory of
liability may be particularly relevant to liability of drivers of early autonomous vehicles. Victims of autonomous
vehicle– related crashes may sue the owners or drivers of the vehicles and argue that the operation of autonomous
vehicle technologies constituted an ultrahazardous activity and the operators should therefore be strictly liable
( regardless of whether they were negligent) for any crashes that occur.
4.1.2 Autonomous Vehicle Technologies and Liability of Drivers
How will autonomous vehicle technologies affect driver liability for automobile crashes? First, these
technologies will likely reduce the number and overall cost of crashes. Human error causes the vast majority of
crashes today, and, by reducing the risk of human error, autonomous vehicle technologies can reduce the incidence
of crashes. This will, in turn, reduce automobile- insurance costs. 16 To encourage adoption, insurers may offer
discounts for operators who purchase automobiles with the appropriate systems. Historically, insurance companies
have been important intermediaries in recognizing the safety benefits of technologies that reduce risk and
( 75 Pa. C. S. § 1705). “ Serious” is defined as “ a personal injury resulting in death, serious impairment of body
function or permanent serious disfigurement” ( 75 Pa. C. S. § 1702).
14 It has proven somewhat disappointing in practice, with costs remaining higher than expected. See
Anderson, Heaton, and Carroll ( in process).
15 This theory of liability is set forth in § § 519– 524A of the Restatement ( Second) of Torts ( 1977): “ One who
carries on an abnormally dangerous activity is subject to liability for harm to the person, land or chattels of another
resulting from the activity, although he has exercised the utmost care to prevent the harm.” Section 520 sets forth the
following factors to be considered in determining whether an activity is abnormally dangerous or ultrahazardous:
( a) existence of a high degree of risk of some harm to the person, land or chattels of others;
( b) likelihood that the harm that results from it will be great; ( c) inability to eliminate the risk by the
exercise of reasonable care; ( d) extent to which the activity is not a matter of common usage;
( e) inappropriateness of the activity to the place where it is carried on; and ( f) extent to which its
value to the community is outweighed by its dangerous attributes.
16 If these technologies reduce crashes sufficiently, it is possible that the very need for specialized
automobile insurance may disappear entirely. Injuries that result from automobile crashes might be covered by
health insurance and homeowner’s liability insurance, in the way that bicycle crashes or other crashes are now
covered. It is not clear how much crashes would have to be reduced in order to make specialized automobile-accident
insurance undesirable. In theory, automobile- accident costs could today be covered under other policies,
though this would require a substantial revision of state law and insurance markets.
2 0
encouraging their policyholders to invest in them. In Europe, for example, insurers have offered a 20- percent
discount on automobile insurance for policyholders who purchase a car with a lane- keeping function and ACC. 17
Second, and more radically, autonomous vehicle technologies may undermine the degree to which a driver
must necessarily be at fault for a crash. Currently, the driver is generally considered exclusively responsible for
control of the vehicle. Hence, we commonly speak of crashes as being caused by one or more at- fault drivers. In the
vast majority of crashes, we ascribe blame to one or more drivers rather than to design features of the car. 18
Autonomous vehicle technologies will likely dilute the sense that drivers are directly and solely responsible for their
automobiles. By shifting responsibility for the automobile from the human driver to the car or its manufacturer,
these systems are likely to undermine the conventional social attribution of blame for automobile crashes. 19
This reduction in fault may be roughly proportional to the extent to which the particular technology
apparently controls the car. As explained in Section 2, the technology operates on a continuum between complete
driver control and complete automobile control. If the technology simply provides additional information to the
driver ( e. g., lane- departure warning), it is less likely to undermine the sense to which the driver retains ultimate
responsibility for the vehicle. In contrast, suppose a car with an autopilot feature engaged crashes into a car in front
of it. If the driver’s use of the autopilot was proper, it seems odd to argue that the driver was necessarily at fault. 20
In crashes that involve drivers reasonably relying on a car’s ability to control itself, there may not be an at- fault
driver for the victim to sue.
While the victims in these circumstances could presumably sue the vehicle manufacturer, products liability
lawsuits are more expensive to bring and take more time to resolve than run- of- the- mill automobile- crash litigation.
This shift in responsibility from the driver to the manufacturer may make no- fault automobile- insurance regimes
more attractive. They are designed to provide compensation to victims relatively quickly, and they do not depend
upon the identification of an “ at- fault” party. 21
Third, this technology may also change the distribution of harms caused by crashes. Presently, the vast
majority of crashes result in relatively minor harm, and these minor crashes vastly outnumber the major ones.
17 It is unclear whether the discount is partially subsidized by the automobile manufacturer ( see Loh, 2008).
18 In theory, injuries suffered in crashes could be said to be caused by the manufacturer’s design decision to
allow automobiles to exceed 20 mph instead of blaming crashes on ( inevitable) driver error. The attribution of
responsibility or causation for a crash ( or any event) is a complex process and that there are a variety of plausible
candidates. Choosing one has policy implications. See Calabresi ( 1975, p. 73).
19 Automobile manufacturers resisted air- bag technology for exactly this reason: concern that legal
responsibility for crashes would shift from the driver to manufacturers ( Wetmore, 2004, p. 391).
20 As tort scholars have long recognized, the assignment of legal responsibility ( and liability) does not, in
theory, need to match our sense of moral responsibility ( see Calabresi and Hirschoff, 1972, arguing that liability
could be placed on the cheapest- cost avoider— the party that is in the position to avoid the costs most cheaply). For
example, one could hold automobile operators strictly liable for crashes that resulted from the operation of their
automobiles, regardless of fault ( defined either morally or economically). More recently, corrective- justice and civil-recourse
tort theorists have questioned whether disconnecting legal responsibility from the sense of wrong is
consistent with many features of tort law. See, e. g., Coleman ( 1992).
21 If one of the proposed forms of universal or near- universal health coverage is enacted, the need for speedy
compensation may become less important.
2 1
Suppose that autonomous vehicle technologies are remarkably effective at virtually eliminating minor crashes
caused by human error. But it may be that the comparatively few crashes that do occur usually result in very serious
injuries or fatalities ( e. g., because autonomous vehicles are operating at much higher speeds or densities). This
change in the distribution of crashes may affect the economics of insuring against them. Actuarially, it is much
easier for an insurance company to calculate the expected costs of somewhat common small crashes than of rarer,
much larger events. This may limit the downward trend in automobile- insurance costs that we would otherwise
expect.
Similarly, new categories of crashes might arise and these might pose interesting questions of how the
liability system sets incentives to coordinate care among parties. Suppose that most cars brake automatically when
they sense a pedestrian in their path. As more cars with this feature come to be on the road, pedestrians may expect
that cars will stop, in the same way that people stick their limbs in elevator doors confident that the door will
automatically reopen. The general level of pedestrian care may decline as people become accustomed to this
common safety feature. But if there were a few models of cars that did not stop in the same way, a new category of
crashes could emerge. In this case, should pedestrians who wrongly assume that a car would automatically stop and
are then injured be able to recover? To allow recovery in this instance would seem to undermine incentives for
pedestrians to take efficient care. On the other hand, allowing the injured pedestrian to recover may encourage the
universal adoption of this safety feature. Since negligence is defined by unreasonableness, the evolving set of
shared assumptions about the operation of the roadways— what counts as “ reasonable”— will determine liability.
Fourth, we think that it is not likely that operators of partially or fully autonomous vehicles will be found
strictly liable with driving such vehicles as an ultrahazardous activity. As explained earlier, these technologies will
be introduced incrementally and will initially serve merely to aid the driver rather than take full control of the
vehicle. This will give the public and courts time to become familiar with the capabilities and limits of the
technology. As a result, it seems unlikely that courts will consider its gradual introduction and use to be
ultrahazardous. On the other hand, this would not be true if a person attempted to operate a car fully autonomously
before the technology adequately matured. Suppose, for example, that a home hobbyist put together his own
autonomous vehicle and attempted to operate it on public roads. Victims of any crashes that resulted may well be
successful in convincing a court to find the operator strictly liable on the grounds that such activity was
ultrahazardous.
Overall, we do not anticipate that liability for individual drivers will be a problematic obstacle to the use of
autonomous vehicle technologies. On the contrary, the decrease in the expected probability of a crash and associated
lower insurance costs that autonomous vehicle technologies will bring about will encourage adoption of these
technologies by drivers and automobile- insurance companies. As responsibility for crashes shifts away from the
driver, no- fault systems may become more attractive.
On the other hand, these technologies pose challenges for manufacturers and may increase their liability
risk in ways that discourage the efficient introduction of these technologies.
2 2
4.2 LIABILITY OF MANUFACTURERS
We anticipate that current liability laws may lead to inefficient delays in manufacturers introducing
autonomous vehicle technologies. The gradual shift in responsibility for automobile operation from the driver to the
vehicle will lead to a similar shift in liability for crashes from the driver to the manufacturer. Recognizing this
effect, manufacturers may be reluctant to introduce technology that will increase their liability.
Liability of automobile manufacturers is governed by product- liability law, which is a hybrid of tort and
contract law concerned with the liability of manufacturers for their products. Substantial variations exist among
states, but many states have adopted portions of the Restatement ( Second) of Torts ( ALI, 1977) and the Restatement
( Third) of Products Liability ( ALI, 1998). These restatements, produced by the American Law Institute, are efforts
to systematize the law. While not automatically binding on any court, many state supreme courts adopt portions of
the restatements to govern particular areas of the law.
Product- liability law can be divided into theories of liability and kinds of defect. Theories of liability
include negligence, misrepresentation, warranty, and strict liability. 22 Types of defect include manufacturing
defects, design defects, and warning defects. A product- liability lawsuit will involve one or more theories of
manufacturer liability attached to a specific allegation of a type of defect. In practice, the legal tests for the theories
of liability often overlap and, depending on the jurisdiction, may be identical. In the next sections, we discuss these
theories of liability, the types of defect, and the way in which each may apply to the categories of autonomous
vehicle technologies.
4.2.1 Theories of Liability
Negligence. Negligence is the most common theory of liability and is the legal standard most often used in
tort law. As explained earlier, the central idea of liability for negligence is that a party should be held liable for
harms it caused by its unreasonably failure to prevent the risk. We explained earlier Judge Hand’s economic
definition of negligence and “ reasonableness.” While equating negligence with cost- benefit analysis has been
criticized by many theorists as reductive and eliminating the subtleties of negligence ( see, e. g., Vandall, 1986), some
form of cost- benefit analysis remains influential in product- liability cases. Recently, courts and the restatement
reporters have been reintroducing the core concept of negligence— reasonableness ( and, in some cases, cost- benefit
analysis)— into other theories of products liability.
While it is difficult to generalize, automobile ( and subsystem) manufacturers may fare well under a
negligence standard that uses a cost- benefit analysis that includes crashes avoided from the use of autonomous
vehicle technologies. Automakers can argue that the overall benefits from the use of a particular technology
outweigh the risks. The number of crashes avoided by the use of these technologies is probably large.
22 We use theory here to mean a legal theory that allows a jury to award damages by connecting a particular
set of facts with legal consequences.
2 3
In contrast, plaintiffs will likely seek to exclude any global cost- benefit analysis that considers the benefits
of avoided crashes and try to focus the reasonableness analysis on the specific facts around the crash ( e. g., was the
automobile manufacturer reasonable in failing to include a warning about sleeping while the ACC and lane- keeping
functions were engaged? What were the costs and benefits of that particular decision?). The plaintiff would likely
argue that evidence about the numerous crashes prevented by this technology was irrelevant. By focusing on the
specific circumstances of the particular crash, plaintiffs will attempt to focus the reasonableness and cost- benefit
analysis away from the long- term safety benefits of these technologies.
Unfortunately, the socially optimal liability rule is unclear. Permitting the defendant to include the long- run
benefits in the cost- benefit analysis may encourage the adoption of technology that can indeed save many lives. On
the other hand, it may shield the manufacturer from liability for shorter- run decisions that were inefficiently
dangerous. Suppose, for example, that a crash- prevention system operates successfully 70 percent of the time but
that, with additional time and work, it could have been designed to operate successfully 90 percent of the time. Then
suppose that a victim is injured in a crash that would have been prevented had the system worked 90 percent of the
time. Assume that the adoption of the 70- percent technology is socially desirable but the adoption of the 90- percent
technology would be even more socially desirable. How should the cost- benefit analysis be conducted? Is the
manufacturer permitted to cite the 70 percent of crashes that were prevented in arguing for the benefits of the
technology? Or should the cost- benefit analysis focus on the manufacturer’s failure to design the product to function
at 90- percent effectiveness? If the latter, the manufacturer might not employ the technology, thereby leading to
many preventable crashes. In calculating the marginal cost of the 90- percent technology, should the manufacturer be
able to count the lives lost in the delay in implementation as compared to possible release of the 70- percent
technology? A host of important definitional issues as to what can be counted as cost and benefit must first be
resolved. These issues make the integration of cost- benefit analysis into tort law complex.
Tortious Misrepresentation. Tortious misrepresentation arises when a party— intentionally or
unintentionally— misleads the plaintiff. There are three types: fraudulent, negligent, and strict liability. Fraudulent
misrepresentation requires that the plaintiff show scienter— actual knowledge on the part of the defendant that the
representation is false. Negligent misrepresentation requires the plaintiff to show that the defendant negligently
asserted some statement as true ( see, e. g., Boos v. Claude, 69 S. D. 254, 1943: used- car retailed advertised car as
being in perfect working order despite defective brakes). In many jurisdictions, this theory of liability has largely
been superseded by the advent of strict liability for misrepresentation. 23 Strict liability for misrepresentation does
23 Section 402B of the Restatement ( Second) of Torts, “ Misrepresentation by Seller of Chattels to
Consumer,” states,
One engaged in the business of selling chattels who, by advertising, labels, or otherwise, makes to the
public a misrepresentation of a material fact concerning the character or quality of a chattel sold by
him is subject to liability for physical harm to a consumer of the chattel caused by justifiable reliance
upon the misrepresentation even though ( a) it is not made fraudulently or negligently and ( b) the
consumer has not bought the chattel from or entered into any contractual relation with the seller.
2 4
not require the plaintiff to prove either fraud or negligence. A false statement of material fact made by the
manufacturer that was justifiably relied on by the consumer can give rise to liability under this theory. 24 However, a
manufacturer will not be found liable for mere “ puffery”— vague positive statements made in advertising about the
product. The line between a material misstatement and mere puffery is not always clear ( see, e. g., Berkebile v.
Brantly Helicopter Corp., 462 Pa. 83, 1975: advertising brochures for helicopter manufacturer did not create liability
where brochures stated that helicopter was “ safe, dependable helicopter” and was “ easy to fly” for beginners,
because these were “ puffs.”).
Tortious misrepresentation may play a role in litigation involving crashes that result from autonomous
vehicle technologies. If advertising overpromises the benefits of these technologies, consumers may misuse them.
Consider the following hypothetical scenario. Suppose that an automaker touts the “ autopilot- like” features of its
ACC and lane- keeping function. In fact, the technologies are intended to be used by an alert driver supervising their
operation. After activating the ACC and lane- keeping function, a consumer assumes that the car is in control and
falls asleep. Due to road resurfacing, the lane- keeping function fails, and the automobile leaves the roadway and
crashes into a tree. The consumer then sues the automaker for tortious misrepresentation based on the advertising
that suggested that the car was able to control itself.
To minimize crashes like this one and to protect themselves from such liability, automakers must be
cautious in promoting autonomous vehicle technologies. This is probably more of a danger as this technology is
initially employed and developed. If a car is, in fact, fully autonomous, this concern disappears because,
presumably, the driver really can rely on the vehicle. But as automakers market technologies that substantially aid a
driver but still require important action on the part of the driver, communicating what the technology can and cannot
do is very important to minimize crashes caused by drivers misunderstanding the capacities of a technology.
Tortious misrepresentation may be a theory used to prevent advertising that leads to these dangerous
misunderstandings.
Warranty. Warranty is a third theory of products liability relevant to autonomous vehicle technologies.
Warranty claims may be divided into express warranties, the implied warranty of merchantability, and implied
warranty of merchantability for a particular purpose.
When a seller expressly warrants some good and the good does not conform to general rules of
merchantability, the buyer may recover. If a consumer buys a toaster, and the toaster shoots flames, the consumer
can recover against the manufacturer. These claims are governed by the Uniform Commercial Code, adopted by
The Restatement ( Third) of Products Liability ( § 9, comment b) adopts the Restatement ( Second) formulation
of misrepresentation: “ The rules governing liability for innocent product misrepresentation are stated in the
Restatement, Second, of Torts § 402B.”
24 See, e. g., Hauter v. Zogarts, 14 Cal. 3d 104, 1975 ( manufacturer is strictly liable for injuries that result
from golf simulator where instruction manual states “ Completely Safe Ball Will Not Hit Player”); compare General
Motors v. Howard, 244 So. 2d 726, 728, Miss., 1971 ( car manufacturer not strictly liable for injuries after steering
wheel advertised as telescopic in a crash did not fully telescope in crash). There have been relatively few reported
cases that cite this section of the Restatement ( Owen, Montgomery, and Davis, 2007, p. 90).
2 5
most states. 25 Victims of automobile- related crashes have recovered under these theories ( see, e. g., McCarty v. E. J.
Korvette, 28 Md. App. 421, 1975: tire guaranteed against blowouts). In Caboni v. General Motors Corp. ( 398 F. 3d
357, 2005), for example, an air bag failed to deploy after a crash. The court explained that the relevant legal question
was whether product performance “ matched that described by the language of the warranty provided with the
vehicle, rather than whether the airbag performed as it was designed to perform.” The scope of the relevant warranty
can be conveyed by text or by illustrations ( see, e. g., Sylvestri v. Warner & Swasey Co., 398 F. 2d 598, 1968:
illustrations showing backhoe used in particular way created express warranty: “ essential idea conveyed by the
advertising representations which is relevant”). Originally, plaintiffs had to prove that they detrimentally relied on
the specific warranty. More recently, however, the courts have moved away from requiring reliance ( Lutz Farms v.
Asgrow Seed Co., 948 F. 2d 645, 1991, citing cases). Some courts have noted the similarity between a theory of
liability based on warranty and § 402B strict liability ( e. g., Klages v. General Ordnance Equipment Corp., 240 Pa.
Super. 356, 1976).
The implied warranty of merchantability suggests that the good sold is generally free from defects and
reasonably fit for the ordinary uses of the good. This is governed by § 2- 314 of the Uniform Commercial Code. 26
While a good does not have to be suitable for all possible uses, it has to be safe for the anticipated uses ( see, e. g.,
Maybank v. S. S. Kresge, 46 N. C. App. 687, 1980: photographic flashbulb that explodes, injuring user’s eye, is not
merchantable). It is closely related to strict tort liability, and most courts equate unmerchantable with defectiveness
under strict- liability standards.
25 Section 2- 313 of the Uniform Commercial Code, “ Express Warranties by Affirmation, Promise,
Description, Sample,” states,
( 1) Express warranties by the seller are created as follows: ( a) Any affirmation of fact or promise
made by the seller to the buyer which relates to the goods and becomes part of the bias of the bargain
creates an express warranty that the goods shall conform to the affirmation or promise. ( b) Any
description of the goods which is made part of the basis of the bargain creates an express warranty
that the goods shall conform to the description. ( c) Any sample or model which is made part of the
basis of the bargain creates an express warranty that the whole of the goods shall conform to the
sample or model. ( 2) It is not necessary to the creation of an express warranty that the seller use
formal words such as “ warrant” or “ guarantee” or that he have a specific intention to make a
warranty, but an affirmation merely of the value of the goods or a statement purporting to be merely
the seller’s opinion or commendation of the goods does not create a warranty.
26 Section 2- 314, “ Implied Warranty: Merchantability; Usage of Trade,” states,
( 1) Unless excluded or modified ( Section 2- 316), a warranty that the goods shall be merchantable is
implied in a contract for their sale if the seller is a merchant with respect to goods of that kind. Under
this section the serving for value of food or drink to be consumed either on the premises or elsewhere
is a sale. ( 2) Goods to be merchantable must be at least such as: ( a) pass without objection in the trade
under the contract description; ( b) in the case of fungible goods, are of fair average quality within the
description; ( c) are fit for the ordinary purposes for which goods of that description are used; ( d) run,
within the variations permitted by the agreement, of even kind, quality and quantity within each unit
and among all units involved; ( e) are adequately contained, packaged, and labeled as the agreement
may require; and ( f) conform to the promise or affirmations of fact made on the container or label if
any. ( 2) Unless excluded or modified ( Section 2- 316) other implied warranties may arise from course
of dealing or usage of trade.
2 6
The last warranty theory of liability is implied warranty for a particular purpose, governed by § 2- 315 of
the Uniform Commercial Code. 27 This provision is applicable when the seller is specifically apprised of the buyer’s
intended use of the good and the buyer relies on the seller’s statements. This theory is less likely to be applicable in
the autonomous vehicle technology context for suits by consumers against manufacturers. It is sometimes used,
however, by manufacturers against part suppliers and might arise in that context ( e. g., Gellenbeck v. Sears, Roebuck
& Co., 59 Mich. App. 339, 1975: supplier of chains for use in swing sets was liable to assembler for sale of chains
that did not resist twisting forces).
In general, sellers can disclaim ( disavow) or limit warranties pursuant to contract law. Some states have
limited merchants’ ability to disclaim warranties.
The warranty theory of liability is unlikely to be particularly important to the development of autonomous
vehicles for two reasons. First, because warranties ( except for the implied warranty of merchantability) can
generally be disclaimed, manufacturers are likely to substantially limit warranties. Second, the implied warranty of
merchantability has merged into strict liability in most jurisdictions, and most of the legal analysis is focused on the
strict- liability theory, which we discuss in the next section.
Strict Products liability. Strict products liability is a fourth theory of liability available to victims in lieu
of having to prove negligence, misrepresentation, or violation of a warranty on the part of the defendant. Strict
liability is conventionally contrasted with negligence because, in theory, it does not require any showing of
negligence, unreasonableness, or any kind of fault on the part of the defendant. In its most extreme interpretation,
strict products liability means that manufacturers insure users against all harms that come from their product,
regardless of fault. Proponents of this loss- spreading rationale argued that strict liability can and should serve this
compensation function and that manufacturers could easily pass on the additional costs of tort judgments to
consumers by raising the prices of their products ( see Priest and Owen, 1985). Manufacturers would also have the
appropriate incentives to reduce the dangerousness of their products. While this expansive version of strict liability
has not been adopted, the rationale of broad liability in order to spread costs underlies strict tort liability as it
actually functions today. 28
In most states, strict tort liability is governed by the Restatement ( Second) of Torts § 402A ( ALI, 1977). 29
Under § 402A, the seller of a product is liable for harm caused by such product when it is sold in a “ defective
27 Section 2- 315, “ Implied Warranty: Fitness for Particular Purpose,” states,
Where the seller at the time of contracting has reason to know any particular purpose for which the
goods are required and that the buyer is relying on the seller’s skill or judgment to select or furnish
suitable goods, there is unless excluded or modified under the next section an implied warranty that
the goods shall be fit for such purpose.
28 Every state except Delaware, Massachusetts, Michigan, North Carolina, and Virginia has adopted strict
liability in tort of some form. The states that have rejected strict liability in tort achieve similar ends through
interpretation of warranty law ( Owen, 2008, § 5.3).
29Section 402A, “ Special Liability of Seller of Product for Physical Harm to User or Consumer,” states,
One who sells any product in a defective condition unreasonably dangerous to the user or consumer
or to his property is subject to liability for physical harm thereby caused to the ultimate user or
2 7
condition unreasonably dangerous to the user.” A product is deemed defective if it “ left the supplier’s control
lacking any element necessary to make it safe for its intended use or possessing any feature that renders it unsafe for
the intended use” ( § 402A). Under § 402A, the imposition of strict liability for a product defect is not affected by the
fact that the manufacturer or other supplier has exercised all possible care. However, litigation about whether a
product is unreasonably dangerous often introduces an analysis of the reasonableness of the manufacturers’
actions— an analysis that usually resembles the analysis of reasonableness that occurs in negligence cases and often
includes one that can include the same cost- benefit analysis issues that were discussed earlier.
Some states have adopted § 2 of Restatement ( Third) of Torts. Published in 1998, approximately 30 years
after § 402A first appeared, it incorporates some of the relevant jurisprudence that has developed as courts have
interpreted § 402A. It retains the traditional strict- liability theory of recovery for claims of manufacturing defect but,
with respect to claims of design defect and inadequate warnings, incorporates an explicit balancing exercise that is
more akin to a negligence analysis. 30
Section 2 adopts a reasonableness- based, risk- utility balancing test as the standard for adjudging the
defectiveness of product designs and warnings. Section 2 also makes clear that even a dangerous product is not
defective unless there is proof of a reasonable alternative design ( see ALI, 1998, § 2, comment d). Comments to § 2
also specify that the risk- benefit balancing done to judge product design must be done in light of knowledge
attainable at the time the product was distributed. The comments also suggest that industry practice and the state of
the art are relevant to the balancing analysis. With regard to warnings, § 2 states that a seller is not liable for failing
to warn of known risks and risk- avoidance measures that should be obvious.
consumer, or to his property, if ( a) the seller is engaged in the business of selling such a product, and
( b) it is expected to and does reach the user or consumer without substantial change in the condition
in which it is sold. ( 2) The rule stated in Subsection ( 1) applies although ( a) the seller has exercised
all possible care in the preparation and sale of his product, and ( b) the user or consumer has not
bought the product from or entered into any contractual relation with the seller. Caveat: The Institute
expresses no opinion as to whether the rules stated in this Section may not apply ( 1) to harm to
persons other than users or consumers; ( 2) to the seller of a product expected to be processed or
otherwise substantially changed before it reaches the user or consumer; or ( 3) to the seller of a
component part of a product to be assembled.
30 Section 2 of Restatement ( Third) of Products Liability states,
A product is defective when, at the time of sale or distribution, it contains a manufacturing defect, is
defective in design, or is defective because of inadequate instructions or warnings. A product: ( 1)
contains a manufacturing defect when the product departs from its intended design even though all
possible care was exercised in the preparation and marketing of the product; ( 2) is defective in design
when the foreseeable risks of harm posed by the product could have been reduced or avoided by the
adoption of a reasonable alternative design by the seller or other distributor, or a predecessor in the
commercial chain of distribution, and the omission of the alternative design renders the product not
reasonably safe; ( 3) is defective because of inadequate instructions or warnings when the foreseeable
risks of harm posed by the product could have been reduced or avoided by the provision of
reasonable instructions or warnings by the seller or other distributor, or a predecessor in the
commercial chain of distribution, and the omission of the instructions or warnings renders the product
not reasonably safe.
2 8
Strict liability under § 402A is probably the theory most often used by plaintiffs in suits involving the
design of automobiles. As such, it will play a central role in litigation over the responsibility for crashes associated
with autonomous vehicle technologies. In the wake of a crash involving an automobile with autonomous vehicle
technologies, victims may argue that the car’s technology was defective in some way. How these principles are
applied will depend, however, on the type of product defect alleged.
4.2.2 Types of Defectiveness
Originally, product- liability case law and doctrine did not distinguish among kinds of defect. Over time,
however, three categories of defect emerged: manufacturing defect, design defect, and failure to warn.
Manufacturing Defects. A product is said to have a manufacturing defect, according to § 2( a) of the
Restatement ( Third) of Products Liability, if it “ departs from its intended design even thought all possible care was
exercised in the preparation and marketing of the product.” Manufacturing defects can be divided into two types.
First, the manufacturer can construct the product using flawed raw materials ( e. g., unduly brittle steel used in a
wheel) ( see, e. g., Magnuson v. Kelsey- Hayes, 844 S. W. 2d 448, 1992). Second, the manufacturer could assemble the
raw materials in an erroneous way— for example, by accidentally severing an important electrical cable during the
manufacturing process. In either case, the product does not meet the manufacturer’s own design specification. 31
If a plaintiff can prove a manufacturing defect by showing that a particular product does not meet the
manufacturer’s own specifications, the manufacturer has very few defenses and is usually found liable. But as “ lean”
production techniques and quality process improvement continues to spread within the automobile industry, the
number of conventional manufacturing defects in automobiles is likely to continue to decrease ( Dassbach, 1994).
Unless a particular autonomous vehicle technology relies on a particular part that is prone to defectiveness
( e. g., a sensor with a high rate of defect), we do not anticipate significant litigation around manufacturing defects. In
cases in which manufacturing defects do occur and lead to crashes, of course, plaintiffs will generally recover.
Instead, more litigation is likely to arise with respect to design defects.
Design Defects. An allegation of a design defect submits that the design of the product itself is defective.
Courts have used two principal tests for defectiveness of design: consumer expectations and cost- benefit. The
consumer- expectation test was explained by one court as follows:
A product is defective in design or formulation when it is more dangerous than an ordinary consumer
would expect when used in an intended or reasonably foreseeable manner. Moreover, the question of
what an ordinary consumer expects in terms of the risks posed by the product is generally one for the
trier of fact. ( Donegal Mutual Insurance v. White Consolidated Industries, 166 Ohio App. 3d 569,
2006)
31 In cases in which the plaintiff cannot allege any particular deviation from the manufacturer’s
specifications, the plaintiff can still ask the jury to infer that an unspecified manufacturing error was responsible for
the accident. For example, in Ducko v. Chrysler Motors Corp. ( 433 Pa. Super. 47, 1994), the court found that a jury
could infer that the car was defective at the time of sale from the fact that it crashed without other explanation, even
absent any specific evidence of defect.
2 9
Comment i to § 402A of the Restatement ( Second) of Torts defines unreasonably dangerous as, in part, an
issue of consumer expectations:
i. Unreasonably Dangerous. The rule stated in this Section applies only where the defective condition
of the product makes it unreasonably dangerous to the user or consumer. Many products cannot
possibly be made entirely safe for all consumption, and any food or drug necessarily involves some
risk of harm, if only from over- consumption. Ordinary sugar is a deadly poison to diabetics, and
castor oil found use under Mussolini as an instrument of torture. That is not what is meant by
“ unreasonably dangerous” in this section. The article sold must be dangerous to an extent beyond that
which would be contemplated by the ordinary consumer who purchases it, with the ordinary
knowledge common to the community as to its characteristics.
While the consumer- expectation test is still used by some jurisdictions, many have abandoned it as
unworkable.
The consumer- expectation test might result in substantial liability for the manufacturers of autonomous
vehicle technologies simply because consumers may have unrealistic expectations about the capabilities of these
technologies ( compare with Hisrich v. Volvo Cars of North America, 226 F. 3d 445, 2000: in product- liability suit
following air- bag deployment leading to fatality, an instruction was warranted on the consumer- expectation test
under Ohio law). Technologies that are engineered to assist the driver may be overly relied on to replace the need for
independent vigilance on the part of the vehicle operator. A definition of design defect that relies primarily on
consumer expectations may result in finding many design defects and lead to substantial manufacturer liability. As
noted earlier ( in discussing tortious misrepresentation), managing consumer expectations to prevent reliance that
exceeds the capacities of these technologies will be important to minimize the number of crashes and reduce
liability.
The cost- benefit ( sometimes also called risk- utility) test is used by many courts to determine whether a
design is defective. It attempts to weigh the benefits, or utility, provided by the particular design against the costs, or
risks, associated with it. In this respect, it is similar to the Learned Hand test for negligence explained earlier.
Many courts have adopted the following factors in conducting the cost- benefit analysis ( Wade, 1973):
• the usefulness and desirability of the product: its utility to the user and to the public as a whole
• the safety aspects of the product: the likelihood that it will cause injury and the probable seriousness of the
injury
• the availability of a substitute product that would meet the same need and not be as unsafe
• the manufacturer’s ability to eliminate the unsafe character of the product without impairing its usefulness
or making it too expensive to maintain its utility
• the user’s ability to avoid danger by the exercise of care in the use of the product
• the user’s anticipated awareness of the dangers inherent in the product and their avoidability because of
general public knowledge of the obvious condition of the product or because of the existence of suitable
warnings or instructions
3 0
• the feasibility of spreading the loss by the manufacturer setting the price of the product or carrying liability
insurance.
The precise factors that courts use in conducting a cost- benefit analysis to determine whether a design is
defective vary by jurisdiction.
Some jurisdictions combine the consumer- expectation and risk- utility tests and allow a plaintiff to prove
design defect by proving either theory ( see, e. g., Barker v. Lull Engineering Co., 20 Cal. 3d 413, 1978). Others have
somewhat confusingly defined consumer expectations in part by a cost- benefit analysis ( Potter v. Chicago
Pneumatic Tool Co., 241 Conn. 199, 1997). Consistent with § 2( b) of the Restatement ( Third) of Products Liability,
many courts require a plaintiff to show that an alternative design was feasible that would have presented the harm.
Current liability law on design defects may hinder the efficient adoption of autonomous vehicle
technologies. Suppose that a particular type of “ autobrake” crash- avoidance technology works to prevent crashes 80
percent of the time. The other 20 percent of the time, however, the technology does not work and the crash occurs as
it would have in the absence of the technology. Victims in those crashes may sue the manufacturer and argue that
the product was defective because it failed to operate properly in their crashes. Under existing liability doctrine, they
have a colorable argument: The product did not work as designed ( manufacturing defect), as advertised ( tortious
misrepresentation), and as warranted ( breach of the implied warrant of merchantability). A manufacturer facing a
decision as to whether to employ such a technology in its vehicles might very well decide not to, purely on the basis
of expected liability costs.
Yet, the social benefits of this technology are vast ( see Parchomovsky and Stein, 2008, arguing that current
tort law inefficiently deters innovation and suggesting reforms). An 80- percent decline ( or even a 10- percent
decline) in even a subcategory of crashes could save many lives and billions of dollars. The existing liability regime
does a poor job of aligning private incentives with the public good in this kind of situation.
One approach to this problem is to integrate a cost- benefit analysis into the standard for liability as
discussed earlier; this integration has the potential to reduce manufacturer liability because it allows the
consideration of the benefits in reduced crash costs that are associated with this technology. If the cost- benefit
analysis is permitted to include these benefits, tort liability for crashes that result from autonomous vehicle
technologies is unlikely because it seems highly likely that the adoption of these technologies is far more likely to
reduce human error and traffic crashes than cause them. In the long run, this may be the socially optimal solution.
But while this may be appropriate calculation of the long- run socially optimal solution, it may also
undermine incentives for safer product design in the short run. Suppose that an auto manufacturer has a choice
between two autonomous vehicle technologies, one of which is much safer but only slightly more expensive. If the
courts adopt a cost- benefit analysis that includes the benefits of the crashes eliminated by an autonomous vehicle
technology, it may be that the manufacturer will not be found liable regardless of whether it chooses the safer or
more dangerous technology. By focusing on the long- run costs and benefits to society of the adoption of
autonomous vehicle technologies, the courts may undermine shorter- run opportunities for efficient safety. In this
3 1
way, conducting the liability cost- benefit analysis by including more costs and benefits may undermine other
incentives for safety in the shorter run. 32
To maximize the social benefits of this technology, policymakers need to structure the liability regime to
encourage the development of this technology without undermining marginal incentives for safety. Further research
is necessary to determine which costs and benefits should be included in the cost- benefit analysis.
We also anticipate litigation around the optimal way to monitor and integrate the driver for those transitional
technologies that are designed to function with a supervisory driver. This issue poses particular difficulties because
alertly supervising the driving function without actively participating may be very difficult for humans, who are
prone to lose attention when not directly engaged. Finding an appropriate way to allow these technologies to
minimize human error without provoking dangerous overreliance on them may prove difficult and prompt post-crash
litigation. 33 It will probably be in automakers’ interest to continue to focus attention on the need for a
responsible driver to monitor the autonomous vehicle technologies, even after the technologies mature sufficiently to
allow truly autonomous operation. To the extent possible, automakers will want to preserve the social norm that
crashes are primarily the moral and legal responsibility of the driver, both to minimize their own liability and to
ensure safety.
Finally, it is also possible that auto manufacturers will be sued for failing to incorporate autonomous
vehicle technologies in their vehicles. While absence of available safety technology is a common basis for design-defect
lawsuits ( e. g., Camacho v. Honda Motor Co., 741 P. 2d 1240, 1987, overturning summary dismissal of suit
alleging that Honda could easily have added crash bars to its motorcycles, which would have prevented the
plaintiff’s leg injuries), this theory has met with little success in the automotive field because manufacturers have
successfully argued that state tort remedies were preempted by federal regulation ( Geier v. American Honda Motor
Co., 529 U. S. 861, 2000, finding that the plaintiff’s claim that the manufacturer was negligent for failing to include
air bags was implicitly preempted by the National Traffic and Motor Vehicle Safety Act). We discuss preemption
and the relationship between regulation and tort in Section 4.3.
Warning Defects. Products can also be found defective for their failure to include appropriate warnings. If
there is a hidden danger in the product, the manufacturer has an obligation to warn of the danger. If it fails to do so,
the product can be found defective as a result of its failure to warn. Section 2( c) Restatement ( Third) of Products
Liability defines a defect for failure to warn as follows:
A product . . . is defective because of inadequate instructions or warnings when the foreseeable risks
of harm posed by the product could have been reduced or avoided by the provision of reasonable
32 See Anderson ( 2007) for a fuller discussion of the way in which the tort system attempts to balances
optimization in the short and long runs.
33 Use of multiple driver- assistance systems at the same time also increases this risk. So, for example,
relying on lane- keeping function and ACC together may pose risks of overreliance that would not be raised if only
one of these technologies was used at a time.
3 2
instructions or warnings . . . and the omission of the instructions or warnings renders the product not
reasonably safe.
There is, however, no clear consensus on an appropriate determination of the conditions in which warnings
are required. Nor is there a clear test as to when existing warnings are adequate. Courts and commentators have
expressed the concern that, after a crash, the cost of an additional warning addressing the circumstances of the
accident will seem infinitesimal to a jury compared with the realized costs of the accident and the injuries suffered
by the plaintiff. As one court explained,
Failure to warn cases have the curious property that, when the episode is examined in hindsight, it
appears as though addition of warnings keyed to a particular accident would be virtually cost free. . . .
The primary cost is, in fact, the increase in time and effort required for the user to grasp the message.
The inclusion of each extra item dilutes the punch of every other item. ( Cotton v. Buckeye Gas
Products Co., 840 F. 2d 937– 939, 1988)
According to this view, courts should be concerned about juries being too quick to find liability for a failure to warn.
There is likely to be substantial litigation over the extent to which warnings are appropriate with
autonomous vehicles. Should a manufacturer warn a consumer that she should not use a laptop computer while
using ACC and the lane- keeping function? The plaintiff in such a case could argue that the cost of such a warning
would be trivial and might save numerous lives. The defendant could argue that it is impossible to anticipate every
situation that could require a warning and that the instruction manual that accompanied the car clearly set out the
limits of the automobile’s autonomous vehicle systems. 34
4.3 EFFECT OF REGULATION AND PREEMPTION
Relevant regulations, engineering standards, and industry custom are usually admissible but not dispositive
as to whether a defendant either met or failed to meet the appropriate standard of care ( Owen, Montgomery, and
Davis