Autonomous vehicle liability: who pays when the car drives itself

Liability for an autonomous vehicle crash can fall on the driver, the manufacturer or the software maker, depending on the level of automation and the jurisdiction. In Britain, the Automated and Electric Vehicles Act 2018 makes the insurer liable first, with a right to recover from the party responsible. In the United States, state law and the courts decide.

EasWrk is a hosting platform, not a research institute. This essay summarises published sources, which are listed at the end. Related reading: AI job displacement for the workforce side, AI in the legal system for how courts treat algorithmic decisions, and the research hub.

The Promise of the Driverless Future

For most of the first two decades of the twenty-first century, autonomous vehicles occupied a singular position in the public imagination of technological progress. They were the canonical example of beneficial artificial intelligence: systems that would eliminate human error from driving, save tens of thousands of lives annually, free up productive time during commutes, extend mobility to the elderly and disabled, and fundamentally reorganize the physical layout of cities freed from the tyranny of parking requirements. The promise was comprehensive, the timelines were specific, and the confidence of the industry's most prominent voices was absolute.

The forecasts were specific. In October 2016 Tesla said all new cars carried hardware for full autonomy, and Elon Musk predicted a coast-to-coast autonomous drive by the end of 2017. In February 2018 the drive had not happened and Tesla said it was still planned (The Drive, 2018). GM said it hoped to start commercializing Cruise vehicles in 2019 (Insurance Journal, 2018). In December 2024 GM stopped funding Cruise's robotaxi work, having invested at least $10 billion in it since 2016 (Fortune, 2024).

The gap between these promises and the actual state of autonomous vehicle deployment in the mid-2020s represents one of the most consequential technology credibility failures in recent memory. Fully autonomous vehicles, those the Society of Automotive Engineers classifies as Level 4 or Level 5, capable of operating without human supervision across a wide range of conditions and environments, remain confined to limited geographic areas, operate under significant constraints, and have a safety record that is actively contested. The technology has advanced significantly; human-supervised autonomous features are now standard on many new vehicles; but the driverless future that was promised as imminent in 2016 has not arrived, and the reasons why illuminate fundamental challenges at the intersection of artificial intelligence, physical systems, and the complexity of the real world.

The SAE automation taxonomy, which classifies vehicles from Level 0 (no automation) through Level 5 (full automation capable of operating in all conditions without human input), provides a useful framework for understanding where the technology actually stands. Most commercially available vehicles in 2026 offer Level 2 features: adaptive cruise control, lane-keeping assistance, and emergency braking systems that can handle specific, well-defined scenarios but require constant human monitoring. Level 3 systems, which can handle driving tasks in defined conditions but require human takeover when prompted, are available from a small number of manufacturers in limited configurations. Level 4 commercial deployment exists in constrained contexts, in specific mapped areas and conditions, primarily from Waymo, which by February 2026 was reported to offer robotaxi service in 10 US metro areas (CleanTechnica, 2026). Level 5 remains a research objective.

The distance between promise and reality matters beyond technology. It affects policy, infrastructure investment, workforce planning, and the legal frameworks that govern liability when automated systems cause harm. Drivers, cities and insurers have all had to plan around forecasts that did not arrive on schedule.

Technical Reality vs. Marketing Narrative

The fundamental challenge facing autonomous vehicle systems is not any single technical problem. It is the accumulated complexity of operating a physical machine safely in an environment that contains an essentially infinite variety of edge cases, unusual conditions, and situations that were not present in the training data. Machine learning systems, including the deep neural networks that power the perception and decision-making components of autonomous vehicles, are extraordinarily capable at tasks for which they have been trained on representative data. They struggle with novel situations that differ meaningfully from their training distribution.

Real-world driving is full of such situations. A mattress on the highway. A traffic cop using hand signals at an intersection where the signals have failed. A school zone with children chasing a ball into the road. Construction that has altered lane markings in ways not reflected in the system's maps. A cyclist behaving unpredictably. An emergency vehicle approaching from a direction not covered by sensor arrays. Human drivers handle these situations through a combination of perception, contextual understanding, common sense, and predictive modeling of other agents' intentions that draws on a lifetime of embodied experience in the world. Current AI systems handle them with varying degrees of success, and the consequences of failure at highway speeds are severe.

NHTSA's Standing General Order, first issued in June 2021 and amended in 2021, 2023 and 2025, requires manufacturers and operators to report certain crashes involving driver assistance systems and automated driving systems (NHTSA, 2025). The reports gave the first systematic public record of such crashes. They count crashes. They do not by themselves give crash rates, because miles driven under each system are not part of the data, so comparing systems with different levels of deployment and operating conditions remains hard.

The 2016 fatal crash in Williston, Florida, in which a Tesla Model S operating with Autopilot engaged did not brake for a semi-truck crossing the highway, killing driver Joshua Brown, became the defining early case study of autonomous vehicle failure (NTSB, 2017). The NTSB adopted its report in September 2017. It gave the probable cause as the truck driver's failure to yield and the car driver's inattention due to over-reliance on vehicle automation. It found that the Autopilot system was not designed to identify a crossing truck or recognize the impending collision, and that the system's design allowed the driver to disengage from driving for long periods and to use it in conditions it was not suited for.

The Waymo v. Uber trade secret case, though a commercial dispute over intellectual property, showed how high the stakes of autonomous vehicle development had become. Uber settled in February 2018 by giving Waymo about 0.34 percent of its equity, valued at roughly $245 million, and agreed not to use Waymo's confidential information (Insurance Journal, 2018). Anthony Levandowski, the former Google engineer at the center of the case, later pleaded guilty to one count of trade secret theft and was sentenced to 18 months in prison (TechCrunch, 2020).

Product names can blur the line between Level 2 driver assistance and Level 4 or 5 autonomy. Tesla's "Full Self-Driving" is a system that requires a supervising driver. In December 2025 a California administrative law judge found that Tesla's marketing of Autopilot and Full Self-Driving was misleading, and the state DMV set a deadline for the company to change it (Reuters via Nasdaq, 2025). The NTSB's Williston findings show why the gap between a name and a system's limits matters: over-reliance on automation was part of the probable cause.

Autonomous Vehicle Accident Liability: Who Is Responsible?

The legal framework governing automobile liability in virtually every jurisdiction in the world is built on the assumption that when a vehicle causes harm, there is a human driver whose negligence or intentional conduct can be assessed and held accountable. This framework, developed over a century of motor vehicle law, insurance regulation, and tort doctrine, is structurally ill-equipped to handle accidents caused by automated systems in which no human was making a real-time driving decision at the moment of the crash.

The fundamental question of autonomous vehicle liability is who bears responsibility when the system makes a decision that causes harm. Is it the vehicle owner, who purchased and deployed the system? The manufacturer, who designed and trained it? The software developer, who wrote the algorithms? The sensor supplier, whose hardware failed to detect an obstacle? The mapping company, whose data was out of date? In a traditional crash, liability analysis focuses on the driver's actions. In an autonomous vehicle crash, the "driver" is a complex sociotechnical system developed by multiple parties, and the decision that led to the crash was made by a machine learning model trained on historical data rather than by a person exercising real-time judgment.

The RAND Corporation's 2016 guide for policymakers (Anderson et al., 2016) devoted chapters to liability. It covered tort liability for drivers and insurers, the liability of manufacturers, and options such as a federal statute limiting tort claims, a no-fault approach, and a presumption that the driver is in control of the vehicle.

Several jurisdictions have developed autonomous vehicle liability frameworks, and the approaches vary. In the United Kingdom, the Automated and Electric Vehicles Act 2018 makes the insurer, or the owner of an uninsured vehicle, liable for damage from an accident caused by an automated vehicle driving itself (section 2), with a right of recovery against the party responsible (section 5). Part 1 of the Act has been in force since 21 April 2021 (legislation.gov.uk). Germany's Act on Autonomous Driving, in force since 28 July 2021, permits Level 4 operation in defined operating areas and sets out duties for the vehicle keeper, a technical supervisor and the manufacturer (German Federal Ministry for Digital and Transport). In the United States, liability has been left mainly to state law and the courts, which creates a patchwork for manufacturers and insurers operating across state lines.

Autonomous car liability under product law

The product liability dimensions of autonomous vehicle crashes are complicated by the nature of machine learning systems. Traditional product liability law distinguishes between manufacturing defects (the specific product deviated from its design) and design defects (the design itself is unreasonably dangerous). A machine learning system that makes a fatal decision in an edge case is arguably neither. It performed as designed, but the training data did not include situations sufficient to handle the scenario it encountered. Whether such a failure fits either category is contested, and this essay found no settled third category in the sources reviewed.

Autonomous vehicle liability insurance

Insurance arrangements are equally unsettled. The shift from human driver liability to product liability for autonomous vehicle crashes implies a shift in who pays insurance premiums, from individual drivers to vehicle manufacturers. But this shift creates complex incentives and information asymmetries. Manufacturers have better information about system capabilities and limitations than insurers do, but they also have financial incentives to minimize the apparent risk profile of their systems. The regulatory infrastructure for transferring liability and actuarial data from the personal auto insurance market to a product liability market does not yet exist at the necessary scale.

Professional Driver Displacement

The workforce implications of autonomous vehicle technology are among its most consequential and least adequately addressed dimensions. Professional driving, including truck driving, taxi and rideshare driving, bus operation, and delivery driving, is one of the largest occupational categories in many developed economies. The Bureau of Labor Statistics counted about 1.98 million heavy and tractor-trailer truck drivers in the United States in May 2022 (BLS, 2022), and the figure for all professional drivers is higher once taxi, rideshare, bus and delivery drivers are added.

Any forecast of driver displacement needs to separate technical potential, what is possible with existing technology, from actual adoption, which depends on cost, regulation, social acceptance and the pace of technical development. Not every technically automatable activity will be automated, and not on the same timeline.

The distinction between technical potential and actual adoption is critical for workforce planning, and it is a distinction that the most alarmist projections about autonomous vehicle displacement have frequently collapsed. Long-haul trucking on interstate highways, which involves relatively predictable driving environments and does not require the final-mile navigation in complex urban environments that is technically most challenging, is the segment of professional driving most likely to see early substantial automation. But even here, the timeline has been repeatedly revised. Aurora began commercial driverless freight runs between Dallas and Houston in spring 2025 (TechCrunch, 2025), a single corridor, so scaled driverless long-haul trucking is still limited to specific routes.

For professional drivers, the uncertainty created by the gap between technological promise and actual deployment has been its own form of harm. Decisions about whether to invest in commercial driver's license training, whether to accept long-term employment contracts in the trucking industry, whether to negotiate for automation transition protections in collective bargaining agreements: all of these decisions have been made in the shadow of a technological transition whose timeline is genuinely uncertain. The workforce harm of technological disruption is not only the disruption itself but the anticipatory uncertainty that prevents rational long-term planning.

Todd Litman of the Victoria Transport Policy Institute concludes that fully automated (Level 5) vehicles will probably become common and affordable in the 2040s to 2060s, and argues that the most optimistic forecasts often come from people with industry financial interests (Litman, 2022). On that view, any workforce displacement would be spread over a longer period than acute disruption narratives suggest.

A more gradual timeline does not eliminate the displacement problem. It may actually make it harder to manage. Gradual displacement without clear milestones is more difficult to anticipate and plan for than sudden disruption. Retraining programs require knowing what jobs displaced workers should be retrained for, which requires knowing what the economy will look like when they complete training, which in turn requires knowing the timeline and character of automation, exactly the uncertainties that make planning difficult.

Self-Driving Car Insurance and the Industry's Identity Crisis

The personal automobile insurance industry, which wrote about $359 billion in direct premiums in the United States in 2024 (NAIC data, reported by Auto Body News, 2025), is built on a century-old model of insuring individual drivers against the consequences of their own and others' negligence. The actuarial models that underpin this industry assume that risk varies meaningfully across individuals, that a 22-year-old male with two speeding tickets represents a different risk profile than a 45-year-old female with a clean record, and that premiums can be priced accordingly. As automation progressively removes the human driver from the risk equation, this actuarial foundation is destabilized.

Swiss Re Institute's sigma 4/2021 report projects that motor insurance will fall from 42 percent of global property and casualty premiums in 2020 to 32 percent in 2040, even as motor premiums rise in dollar terms, with technology a main reason for slower growth (Swiss Re Institute, 2021). The report does not set out the three scenarios below. They are this essay's illustrations of how the market could change. In the first, a slow and partial adoption scenario, personal auto insurance evolves gradually as automated features reduce but do not eliminate human driver risk, with insurers developing new rating factors around the mix of automated and manual driving miles. In the second, a rapid adoption scenario, the shift from personal to commercial/product liability creates a fundamental restructuring of the insurance market, with the personal auto line shrinking substantially as fleet operators and manufacturers become the primary insurance purchasers. In the third, a mixed scenario reflecting the reality of uneven adoption across geographies and demographics, both markets coexist for an extended period, creating complex underwriting challenges.

One data challenge follows from the scenarios: as vehicles spend more time in automated mode, the human-driven miles that insurers have traditionally used to price risk become less representative of actual risk exposure. Take a hypothetical vehicle that drives 80 percent of its miles autonomously and 20 percent manually. It creates an actuarial problem: the human-driven portion may be disproportionately concentrated in the complex urban and adverse-weather scenarios that automated systems handle least well, making the remaining manual driving more risky than the raw mileage numbers suggest.

The insurance implications extend to commercial and liability markets. If autonomous vehicle crashes become primarily product liability events, the commercial insurance arrangements needed to support that liability are fundamentally different from personal auto insurance. Product liability policies for a manufacturer whose vehicles travel billions of miles annually would need to be of an entirely different scale and structure than anything in the current commercial insurance market. The actuarial data required to price such policies does not yet exist at the necessary volume or granularity.

The core challenge for insurers is pricing risk for a technology whose safety record is still being established, in a liability environment that is legally unsettled. That makes caution the likely stance. It also helps explain why the question of self-driving car insurance cost has no settled answer yet.

Urban Planning and Infrastructure Implications

The anticipated arrival of autonomous vehicles generated a wave of urban planning speculation in the 2010s, much of which now needs revision in light of the actual adoption trajectory. Many planning scenarios assumed rapid adoption of shared autonomous vehicle fleets that would reduce private vehicle ownership, free up parking, and change land use around streets that no longer need wide lanes.

Fagnant and Kockelman's simulation work estimated that each shared autonomous vehicle could replace around 11 conventional vehicles in the modeled scenario, with large implications for ownership, parking demand and land use (Fagnant & Kockelman, 2014). Their model also counted extra travel from vehicles repositioning empty. These are modeled results, not observed outcomes.

Whether shared autonomous fleets will replace private cars at modeled rates has not been shown at scale. In San Francisco, where Waymo and Cruise both ran commercial robotaxi services, an October 2023 crash in which a Cruise vehicle dragged a pedestrian led California regulators to suspend Cruise's permits on 24 October 2023 (TechCrunch, 2023), and Cruise paused all US operations days later.

Infrastructure requirements for autonomous vehicles have also proven more demanding than early analyses suggested. Many autonomous systems use high-definition maps for precise localization, and those maps need continuous updating as roads, construction and signage change. Maintaining them across a whole city is likely to be costly, and map dependence is a weakness in construction zones, temporary detours and unfamiliar places.

The Road Ahead

The trajectory of autonomous vehicle technology in the coming decade will be shaped by the resolution of three fundamental uncertainties: the pace of technical progress on the edge cases that have resisted solution; the evolution of the regulatory and liability frameworks that determine where and how autonomous vehicles can operate commercially; and the economic viability of autonomous vehicle services at the margins of geographic and weather conditions where they currently cannot operate.

The most likely near-term trajectory is continued expansion of limited-area autonomous operations alongside the gradual refinement of advanced driver assistance systems in mass-market vehicles. Waymo has already expanded to additional cities (CleanTechnica, 2026). Tesla's approach of collecting training data from its consumer fleet and improving its driver assistance through over-the-air updates continues, and California regulators have already scrutinized its marketing.

The workforce transition challenge will not be resolved by any single policy intervention. It requires a combination of realistic timeline projections that avoid both alarmism and false reassurance; investment in social insurance mechanisms that provide income support during transitions; and portable benefits systems that reduce the dependence of workers' security on continued employment in any specific occupation. The Fagnant and Kockelman vision of cities reorganized around shared autonomous mobility remains a possible long-term future, but a horizon measured in decades, and the workers and communities in its path deserve planning horizons that match.

The liability and insurance questions will require active legislative and regulatory engagement in most jurisdictions. Market mechanisms alone will not produce a coherent liability framework for autonomous vehicle crashes, because the information asymmetries, multi-party causation questions, and novel legal categories involved are beyond what common law evolution can efficiently resolve. The UK's approach of establishing clear initial insurer liability with manufacturer recovery rights provides one model worth examining, though its adequacy at scale remains to be tested.

The driverless future has not failed. It has been deferred, and its deferral has revealed the complexity that oversimplified early projections obscured. The social, legal, and economic infrastructure required to support widespread autonomous vehicle deployment is at least as consequential as the technical infrastructure, and it has received a fraction of the investment. Closing that gap is the work of the coming decade.