Who Bears the Blame When a Self-Driving Car Crashes — The Boundary Between Product Liability and Driver Liability

Why This Debate?

On March 23, 2018, a Tesla Model X crashed into a highway barrier in Mountain View, California. The driver was Walter Huang (38 at the time), an Apple engineer, and he died. Immediately after the crash, Tesla denied responsibility. Tesla's position is this. "The fundamental premise of moral and legal responsibility is a broken promise, and there is no such thing here. He knew well that Autopilot was not perfect, and this accident could only happen when he was not paying attention to the road even though the vehicle gave multiple warnings" (DOI: 10.31228/osf.io/ur3e5). Autopilot is Tesla's automated driving feature. In other words, the system was not perfect and the driver knew it. The logic is that there would have been no accident if only he had paid attention. This is the first front surrounding autonomous driving accidents.

An uncomfortable question is hidden here. To someone who was not holding the steering wheel, how far can we say, "It is your fault"? Negligence is the wrong of failing to exercise the care that one reasonably should have exercised to prevent an accident. In a situation where the system drives and a human only intervenes occasionally, how justified is the demand that "you should have paid attention"?

First, we clarify which country's law this article deals with. This is not an article explaining the law of a single country. The place where the accident happened is the United States. The laws dealt with by the papers cited vary by country. There are the United States, the EU (Germany, France, the Netherlands), the UK, Canada, and Singapore. There are also China, South Korea, Japan, Indonesia, Kuwait, and Vietnam. Therefore, whenever a court or law appears, we note separately which country's story it is.

Some countries have already changed their laws. However, the broad framework of how and to whom to divide responsibility differs by country and has not been agreed upon. Much of what this article deals with is not established case law, but proposals and arguments by scholars and lawmakers in each country. Papers uploaded to SSRN sometimes appear as preprints before peer review. Such papers are marked as "not peer-reviewed". This article does not judge which side is right. The purpose is to look squarely at that unsettled state without smoothing it over.

Common Belief: The Driver is Responsible When an Accident Happens

Traditionally, responsibility on the road has been explained around the driver's negligence. The common belief is that if the driver has fault, that person pays compensation. Countries also place liability on vehicle operators or presume negligence. In the negligence-centered explanation, the law examines who violated the signal, who failed to look ahead. It also examines who failed to fulfill the duty of care. The duty of care is the caution one must exercise not to harm others. Under this liability for negligence, the victim must directly prove the perpetrator's intent or negligence. That proof was the gateway to compensation (DOI: 10.58837/lawchulajournal.v44i1.281356).

There is a reason this common belief gained strength. It is because it is commonly said that the vast majority of fatal accidents stem from human error. There is an expectation that self-driving cars will save tens of thousands of lives every year. This expectation also assumes the premise that "the overwhelming majority of fatal accidents are human error" (DOI: 10.1515/jtl-2019-0029). If the driver is the origin point of the accident, holding the driver responsible is natural. The common belief starts right here.

The Twist: The Moment You Let Go of the Steering Wheel, the Principle Collapses

The problem is that autonomous driving shatters that very premise—namely, the premise that "the human is the one driving." In highly automated driving, driving is no longer in the hands and feet of a single person. Software, sensors, updates, communications, and infrastructure all participate in driving. Sometimes remote supervision gets involved as well. Driving intertwined among multiple entities like this is called a distributed socio-technical process (DOI: 10.31743/recl.19473). The traditional responsibility model centered on the driver fundamentally conflicts with this structure. There is an analysis that this mismatch creates systemic gaps when dividing responsibility (DOI: 10.51204/anali_pfbu_26201a).

We will point this out now because it is a phrase that will appear often. Autonomous driving is divided into six stages, from the stage where a human does all the driving to full automation. This is a classification by the Society of Automotive Engineers (SAE), and the stages are called levels.

Let us return to the Tesla incident mentioned earlier. Scholars who study the interaction between humans and automation in aviation and military fields directly refute Tesla's logic. Their argument goes like this. Science plays a decisive role in dividing legal responsibility. Courts rely on scientists and engineers when determining whether an accident could have been foreseen beforehand—that is, whether it was foreseeable. If science knew about a certain relationship at the time of the accident, the designer can be held responsible. That relationship is the systematic relationship between the accident and the preventive measures the designer failed to take. The conclusion of decades of research is this: A design that automates everything that can be automated and dumps the rest onto humans is inappropriate and dangerous. According to them, there is also a systematic relationship between the design of autonomous vehicles and the types of accidents happening now. Therefore, these accidents were not unforeseeable and were not solely the driver's fault (DOI: 10.31228/osf.io/ur3e5). Of course, this is the argument of these authors. How courts will judge each case is a separate matter.

When conventional wisdom is shaken, gazes turn beyond the driver. There are many candidates for responsibility. Drivers, owners, manufacturers, software suppliers, telecommunication network operators, component makers, and even governments—anyone can be a candidate. There is an analysis that traditional tort law rules alone cannot sort out which of them should be held accountable (DOI: 10.3390/wevj12020062). Tort law is the law that requires compensation for harm caused to others even among those without a contractual relationship. When a system's judgment, rather than a human's mistake, causes an accident, accident law loses the starting point where it used to find the cause in the human.

There is an analysis comparing the civil law causation theories of Indonesia and Kuwait. Civil law is the law dealing with rights and obligations between individuals, and among them, the rules on damages are the concern of this article. Causation judgment is the task of sorting out whose action produced the damage. This analysis views that task as being pushed out of traffic law into general civil law and the realm of product defects (DOI: 10.47268/ballrev.v7i1.3700).

Mechanism: Where to Shift the Responsibility

If the driver is not the only answer, what are the alternatives? This is where the logic of various factions diverges. Each faction brings out different legal tools, and those tools often clash. The seven branches below are competing proposals, not established answers.

Camp 1 — Product Liability: "Prove the Defect"

The most prominent alternative is to view autonomous driving accidents as a matter of product liability. Product liability is the responsibility of the maker to compensate for damages caused by a product's defects. If a car drives itself, the idea is that an accident should be seen as a product defect rather than the driver's mistake.

Let us clear up a common misconception here. It is the belief that product liability is entirely strict liability. Strict liability is responsibility borne regardless of whether one exercised due care. A paper that has not undergone review concludes that the actual structure is more subtle. Although the abstract does not specify the country, this distinction appears to presuppose American product liability jurisprudence. Manufacturing defects carry strict liability regardless of whether due care was taken. Design defects and warning defects are judged within a negligence framework based on risk-utility analysis (DOI: 10.2139/ssrn.6687638). A manufacturing defect is a mistake made during the production process. A design defect is a flaw in the design itself. A warning defect is a flaw stemming from inadequate warnings or instructions. Risk-utility analysis is weighing a product's benefits against its risks. Even if a product's utility is enough to justify its risks, liability does not vanish. The focus simply shifts to whether the manufacturer took reasonable steps to reduce foreseeable risks through warnings or instructions. The EU Product Liability Directive uses a strict liability framework where design and warning defects are also judged by whether a defect exists, not by negligence. Therefore, this distinction does not transfer over as is.

In autonomous driving, all three of these defects become problematic. One must determine whether the failure in design, manufacture, or warning was a "sufficient cause" of the victim's damages (the preceding comparative analysis of Indonesia and Kuwait, DOI: 10.47268/ballrev.v7i1.3700). A particularly tricky issue is the failure of human-machine interaction. As automation progresses, the interaction between the driver and the system changes fundamentally. What if an accident occurs because that interaction was not smooth? Can this breakdown itself be considered a "defect" under the EU Product Liability Directive (85/374/EEC)? One paper poses this question (DOI: 10.1093/ijlit/eaac010).

The strong argument from the product liability camp is this. The fact that autonomous vehicles are statistically safer should not automatically serve as grounds for exemption from liability. This paper finds its basis in the history of product liability lawsuits (the abstract does not specify the country). The litigation system generally did not hold a car liable simply because it was an old, dangerous car lacking the latest safety features. Nor did it exempt a new, safe car simply because of its superior aggregate performance. Following the same logic, this author argues: Just because a conventional car is not a defective car despite relying on a human driver, an autonomous car should not be exempted from liability either if its defect caused harm (DOI: 10.1515/jtl-2019-0029).

Camp 2 — No-Fault and Strict Liability: "Give Up on Finding Negligence"

The second camp is more radical. They suggest ditching the negligence framework of asking "whose fault is it?" entirely. Strict liability is often introduced as an exception to fault-based liability. Unless vetted, one theoretical analysis overturns this. According to this analysis, strict liability is not an exception outside the logic of negligence. It is an institutional decision to allocate a specific category of risk to the actor from the start. Abnormally dangerous activities, product liability, and vicarious liability all share this structure. Vicarious liability is the responsibility a person who employs others bears for the actions of those they employ. From this perspective, negligence is not the foundation of responsibility, but merely one of several pathways through which responsibility arises (DOI: 10.2139/ssrn.6168207).

A related argument is the theory of enterprise liability. One paper views AI as complex and beyond human control. Therefore, demanding that a victim "pinpoint whose negligence it is" becomes an unfair burden. It also states that it is extremely difficult to clearly prove the causality between the AI's behavior and the damage. This paper suggests that manufacturers should be held responsible for defects regardless of negligence. The victim's burden is reduced to proving that "the AI had a defect and that defect caused the damage." This model views responsibility as a matter of risk distribution and internalizing costs. It encourages manufacturers to calculate the cost of potential damages in advance and distribute it to the market (DOI: 10.58837/lawchulajournal.v44i1.281356).

Camp 3 — Reinterpreting Negligence: Redrawing the Standard of Foreseeability

The third camp suggests redrawing the concept of negligence instead of discarding it. The core is foreseeability, meaning whether the outcome could have been anticipated in advance. This is a primary standard courts use when determining negligence. AI developers often defend themselves by saying, "The outcome could not be foreseen. The causal pathway was novel, complex, and opaque."

A paper analyzing how U.S. courts apply foreseeability argues that this defense misallocates responsibility. In this paper, opacity and unpredictability are not intrinsic properties of advanced AI. They are the result of developers' abstraction choices that prioritize accuracy and efficiency over interpretability. If such choices increased the possibility of opaque outcomes, the law should also be adjusted to reflect that responsibility, it argues. Specifically, the proposal is to relax the foreseeability standard—which is disadvantageous to the plaintiff (the victim who filed the lawsuit)—into a more favorable direction. In particular, it views the stage of the duty of care as the critical point for reform. However, it suggests maintaining factual adjudication at the stages of negligence determination and causation. This is to prevent responsibility from becoming excessively broad (DOI: 10.1515/jtl-2025-0027).

Another way to reconstruct negligence is to establish a legal fiction called the "computer driver." A legal fiction is a device where something is not actually a person, but is agreed to be treated like one in the law. This proposal came out in the form of a legislative bill. A court or jury judges whether the computer driver exercised as much care as a "careful and fully competent human driver." The idea is that if it is judged to have fallen short, fault-based liability is imposed on that computer driver (DOI: 10.25172/smustlr.27.1.5). The system placed in the position of a human will be judged by the negligence standards previously applied to humans.

Faction 4 — The Moment of Handing Over: The Level 3 Gray Area

The trickiest zone to handle is Level 3 (Conditional Automation). This is the stage where the system drives and hands control back to the human when necessary. The system's request for the human to take back control is called a takeover request. The problem is that this moment of handoff can become a vacuum of responsibility. The issues of illegal acts, product liability, and privacy in Level 3 vehicles are treated as a separate area of discussion. This is a discussion attempting to apply existing legal principles to this narrow area (DOI: 10.36645/mtlr.30.2.tort).

The computer driver proposal from earlier tackles this gray area head-on. The time during which control passes from the computer to the human is called the transition period. The point is to nail down that human drivers cannot legally bear contributory or comparative negligence during a certain section of this transition period (DOI: 10.25172/smustlr.27.1.5). Contributory negligence and comparative negligence are legal principles where compensation is reduced or eliminated if the victim also shares some blame.

Apart from this proposal, there are human factors experiments measuring human reactions at the moment of handoff. This is a small-scale Level 3 simulator study involving 18 drivers. Takeover reaction times were longer in urban scenarios than on highways. As the burden of non-driving tasks like music, video, and news summaries increased, takeover performance tended to worsen. In particular, the news summary task delayed reactions the most and reduced situational awareness (DOI: 10.54941/ahfe1007862). This study speaks to the need for handover support design, not legal responsibility. As we see it, this result suggests that the demand of "we gave a warning so you should have responded immediately" may conflict with human cognition. However, because it is a small-scale simulator study, it is premature to use it as a basis for legal judgment.

Camp 5 — Criminal Liability: Putting the Manufacturer in the Dock

Moving beyond civil damages to criminal liability (criminal liability) makes the debate much hotter. Criminal liability is responsibility for which the state punishes a crime. If civil law is a matter of patching up harm with compensation, criminal law is a matter of punishment. This discussion often takes place on the premise that the autonomous driving system itself cannot be the subject of criminal liability. If so, who bears the criminal liability for an accident caused by the system?

In Chinese criminal law discussions, there is a paper that examines the manufacturer's duty of care throughout the entire process from production to application. This paper argues that manufacturers can be held responsible through the product infringement theory. This paper concludes that it is reasonable to view the manufacturer's criminal liability as the crime of traffic accidents. Furthermore, it writes that it hopes to add autonomous vehicle manufacturers as a unit subject of the crime of traffic accidents, namely as an organization like a company (DOI: 10.54097/vswsyq46). However, this is a legislative proposal, not current law.

Applying criminal negligence to autonomous driving runs into conceptual difficulties. A comparative law study calls this "negligence failures". It counts epistemic problems, a lack of control, and generalized risk as causes. To these is added the "problem of many hands". The problem of many hands is when multiple actors are tangled up, blurring whose responsibility it is. Such problems are said to shake the components of criminal negligence, which are awareness, foreseeability, and risk-taking. Singapore, France, and the UK are responding with different models. Recommendations include AI training and licensing systems for users, and empirical verification of handover requests (DOI: 10.1007/s10609-023-09451-1). There is also criticism that at the EU level, only technical standards are being pushed while regulations on criminal responsibility are missing. In EU criminal law discussions, there is also a proposal to distribute criminal liability for "when humans put AI in the driver's seat" (DOI: 10.1177/20322844231213336).

Germany's Volkswagen case was handled not within criminal law, but within the civil product liability framework (Product Liability Act, class action suits, compensation funds). It can only be read as a mirror reflecting the difficulty of criminal liability, and is not a criminal precedent. There is a study analyzing Volkswagen's AI-assisted driving functions. This study looks at it alongside the German Road Traffic Act (StVG) and Product Liability Act (ProdHaftG). StVG §63a is a regulation stipulating data recording during automated driving. It looks at complex accidents involving machine learning. It points out that when the judgment of an opaque AI model is the cause, it is difficult to prove causality.

The study analyzes a 2023 Braunschweig Regional Court ruling. It also covers a lawsuit pending in the Berlin Court of Appeal at the time of the 2025 paper. The study's evaluation is that this Volkswagen case reveals a recurring vacuum. The question is whether collision simulation data can be used as evidence, in other words, its admissibility as evidence. There is also no standardized documentation on software updates and system structures. This article could not confirm the contents of the ruling and subsequent developments. This study's proposal is regulatory innovation. Examples include mandatory "software liability documentation", standards for the admissibility of simulation evidence, and a compensation fund dedicated to autonomous vehicles (DOI: 10.55843/icl2025cong223p).

Camp 6 — Data and Black Boxes: Evidence Divides Responsibility

At the bottom of all the preceding debates lies a single practical problem: How do you prove what happened? That is why car black boxes, or event data recorders, are emerging as core evidence for judging responsibility. There is a study analyzing China's Xu case and technical standards (GB44497-2024). The analysis shows that black box data is decisive evidence, but has limitations in extreme weather conditions. The same analysis also highlights the L1 to L5 automation grading system. It suggests this system can be used to clarify responsibility among manufacturers, software suppliers, and drivers (DOI: 10.54254/2753-7048/2025.bo23203).

Black boxes create new problems of their own. In vehicles where software is frequently updated, update paths can become channels for malicious tampering. A trustworthy party might also accidentally install the wrong version. That is why a dececentralized black box audit trail design combining distributed hash tables and public blockchains is proposed (DOI: 10.4271/11-03-02-0006). An audit trail is a record leaving who changed what and when. The goal is to guarantee the integrity of vehicle software—meaning that records have not been altered—and its origin.

A more fundamental barrier is data access rights. Information from black boxes and event data recorders is usually held by insurance companies and manufacturers. Legal hurdles like privacy, intellectual property, and competition block their reuse. This is why Europe's Data Act (which was at the proposal stage at the time of the paper) is examined as a solution to this access rights problem (DOI: 10.12688/openreseurope.16468.1).

The difficulty of evidence directly leads to the problem of the burden of proof, meaning who has to prove it. The EU's proposed AI Liability Directive (AILD) included strengthened rules for evidence disclosure and the presumption of causality. It aimed to ease the victim's burden of proof. At the time this analysis was written, it was a proposal-stage draft. The Commission announced plans to withdraw this proposal in February 2025. Evidence disclosure is a procedure that forces the opposing party in a lawsuit to hand over evidence they hold. A presumption of causality is a device that provisionally treats something as the cause even if the victim does not prove the entire causal relationship. However, this analysis, which deals mainly with generative AI, argues that even that is insufficient. The claim is that the risks inherent in AI itself can be the basis for aligning EU rules toward strict liability (DOI: 10.1093/ijlit/eaae021).

Camp 7 — Insurer-First: "Compensate First, Argue Later"

Some camps suggest pushing the tedious debate over who is responsible aside altogether. This is the insurer-first compensation model. It aids victims quickly first. The idea is to sort out the allocation of responsibility—whether it lies with the manufacturer or the driver—later through subrogation. Subrogation is the right of the party that paid compensation first to demand that the money be returned from the party actually responsible.

A paper examining EU law observes that the EU does not rely on a single autonomous driving liability code. Instead, EU law combines three things. First is the insurer-first logic of victim compensation. Second is the modernization of product liability targeting software damages. Third is risk-based regulation that imposes documentation and post-market obligations on front-end actors (upstream actors) (DOI: 10.31743/recl.19473).

Countries are tinkering with their existing compulsory insurance frameworks in this direction. A paper by Korean academia summarizes the amendments in Germany and Japan as follows. Germany doubled the liability cap for autonomous vehicle accidents through an amendment to the Road Traffic Act (StVG). It also introduced the concepts of Level 4 and 5 and the status of a technical supervisor (a person who technically supervises the vehicle). Japan maintained its existing liability system. It established new regulations targeting driverless driving and a licensing system to be used for regional mobility services. Discussions to overhaul the Automobile Damage Compensation Guarantee Act are also underway in Korea. They start from the critical awareness of quickly aiding victims and lowering social costs in the era of autonomous driving (DOI: 10.18189/isicu.2024.31.2.149). The United States has laws scattered across states. A 2022 survey identified related legislation in 30 states. However, it points out that data to evaluate its effectiveness is still lacking (DOI: 10.1017/jme.2022.96).

Debate and Unresolved Issues: No Grand Framework is Agreed Upon

We have surveyed seven camps so far, but to be frank, none of them has won. The reality is that multiple frameworks are mixed and competing. One study comparing the liability systems of autonomous driving accidents organizes this into four types. They are driver liability, system liability, manufacturer-operator liability, and a complex liability mixing these. It argues that unique institutional dilemmas are revealed in each type. The boundaries of responsibility are vague. Mechanisms for securing evidence and acknowledging facts are inadequate. Responsibility distribution is unbalanced and judgment rules are fragmented (DOI: 10.3390/wevj17010032).

A study analyzing China's legal and industrial environment offers a compromise to give different answers depending on the automation level. Levels 0 to 2 place responsibility on the driver. Level 3 divides it between the driver and the manufacturer or operator depending on the road and system conditions. Levels 4 to 5 propose placing it mainly on the manufacturer, operator, and system supplier (DOI: 10.3390/wevj16120665). This is a proposal premised on China and is not an internationally organized standard.

An essay comparing the UK, the US, and Canada offers a mixed standard containing a rebuttable presumption. This essay has not been peer-reviewed and is written by an undergraduate author. A rebuttable presumption is an assumption that is taken to be true for now but can be overturned if contrary evidence emerges. Under this standard, design, warning, and failure during automated driving are in principle presumed to be the manufacturer's responsibility. System misuse or unreasonable failure to respond to a takeover request is attributed to the operator's responsibility. Furthermore, it suggests shifting the burden of proof using event data logs, and having insurers compensate first and then seek reimbursement (DOI: 10.2139/ssrn.6543563). It has gathered tools from different camps into a single framework, but this too is merely one proposal.

Before all these sophisticated proposals, there are also voices telling us to take a step back. One paper argues as follows. It is clear that autonomous vehicles will significantly change the practice of car accident liability systems. However, how they will change is not yet known (DOI: 10.1515/jtl-2020-2012). Attempts to imagine the contours of the future and redesign the liability system in advance are prone to causing errors and frustration. Therefore, the argument is to muddle through for now, meaning to try things out and fix them as we go. This is an attitude of being cautious instead of rushing to judgment in the face of an unsettled problem.

National discrepancies also fuel this undecided state. The previous EU paper uses Germany, France, and the Netherlands as example models and applies Vietnamese law to the same framework. According to this paper, Vietnam already has a foundation for victim protection through strict no-fault legal doctrines like "abnormally dangerous sources." However, it states that risks of persistent mismatching still remain in Vietnam regarding access to evidence, attribution of cyber accidents, and coordination of criminal liability (DOI: 10.31743/recl.19473. In short, the proposals examined in this piece are not converging into a single correct answer, but are splitting into multiple paths and being experimented with.

Significance: Technology is rewriting the legal concept of 'negligence'

Then what is the true meaning of this debate. On the surface, it looks like a practical problem of "who will pay compensation for an autonomous vehicle accident." But what is happening underneath is much larger. Technological progress is pressing the fundamental legal concept of 'negligence' to redefine itself.

For over a century, car accident law has been structured around "finding the faulty human and holding them accountable." Autonomous driving removes the driving human, the heart of that structure. Different camps answer that empty space differently. Product liability says, "prove the defect." Strict and enterprise liability say, "give up finding fault and distribute the risk." The negligence reinterpretation camp says, "let's redraw foreseeability." Criminal law asks, "can manufacturers be put in the defendant's seat?" The insurance-first camp answers, "compensate first before arguing." All these proposals are ultimately variations of a single question expressed differently. In an accident where nobody held the steering wheel, what is responsibility.

Let us reiterate that most of what this article covered are proposals and arguments, not established precedents. So what we are seeing now is not the discovery of a correct answer, but the process of concepts being rewritten. The next time you hear the question "who is responsible if an autonomous vehicle causes an accident," it would be good to keep this point in mind. There is still no single answer to that question. The absence of that answer can be read as a signal that the law is rewriting itself to catch up to technology. The scales of negligence are being remade to fit the age where we let go of the steering wheel.

Source Papers

• "Unsafe at Any Level: The U.S. NHTSA's Levels of Automation Are a Liability for Automated Vehicles," Communications of the ACM (2020) — DOI: 10.31228/osf.io/ur3e5 — The Walter Huang/Tesla case, Tesla's logic in denying liability and the counterarguments, and the claim that the crash was foreseeable and not solely the driver's fault.
• "Autonomous vehicle and the limits of traditional liability," Anali Pravnog fakulteta u Beogradu (2026) — DOI: 10.51204/anali_pfbu_26201a — The driver-centric traditional liability model structurally conflicts with high automation, creating systematic gaps.
• "Attributing Liability for Autonomous Vehicles: EU Multi-Level Approaches and Implications for Vietnamese Law," Review of European and Comparative Law (2026) — DOI: 10.31743/recl.19473 — Viewing driving as a distributed socio-technical process, combining the EU's insurance-first approach, modernization of product liability, and risk-based regulation.
• "Rethinking Producer Liability in Autonomous Vehicle Accidents," Batulis Civil Law Review (2026) — DOI: 10.47268/ballrev.v7i1.3700 — Comparing causation theories in Indonesian and Kuwaiti civil law. Human-driver-centric norms are no longer adequate, and causation judgments are shifting to the domain of product defects.
• "Who Is Liable When a Driverless Car Crashes?" World Electric Vehicle Journal (2021) — DOI: 10.3390/wevj12020062 — Traditional tort rules fail to single out the multiple liable candidates, making product liability the core issue.
• "Reframing Product Liability: Beyond Strict Liability," (2026) — DOI: 10.2139/ssrn.6687638 — The dual structure of product liability: manufacturing defects equal strict liability, while design and warning defects follow a risk-utility-based negligence framework.
• "Autonomous Vehicles, Technological Progress, and the Scope Problem in Products Liability," Journal of Tort Law (2019) — DOI: 10.1515/jtl-2019-0029 — Relative safety cannot serve as an exemption ground, and defective self-driving cars are not immune from liability.
• "Human–machine interaction in self-driving vehicles: a perspective on product liability," International Journal of Law and Information Technology (2022) — DOI: 10.1093/ijlit/eaac010 — Whether the breakdown of human-machine interaction can be regarded as a 'defect' under the EU Product Liability Directive.
• "Liability for Autonomous Vehicle Torts: Who Should Be Held Responsible?" World Electric Vehicle Journal (2025) — DOI: 10.3390/wevj16120665 — Analyzing China's legal and industrial environment. Liability allocation by automation level (L0–2 drivers, L3 shared, L4–5 manufacturers, operators, and systems).
• "Strict Liability and the Geometry of Risk," (2026) — DOI: 10.2139/ssrn.6168207 — Strict liability is not an exception to negligence, but an institutional decision to assign risk to the actor from the outset.
• "Why AI Liability should shift from fault-based theory to Enterprise Liability," Chulalongkorn Law Journal (2026) — DOI: 10.58837/lawchulajournal.v44i1.281356 — A proposal to ease the burden of proving negligence and impose no-fault, risk-distributing enterprise liability on manufacturers.
• "Foreseeing the Unforeseeable," Journal of Tort Law (2026) — DOI: 10.1515/jtl-2025-0027 — Analyzing U.S. tort law. Opaque nature is the result of developers' abstraction choices; let us relax the foreseeability standard in favor of plaintiffs.
• "Tort Liability, Privacy, and Regulatory Considerations for Level 3 Autonomous Vehicles," Michigan Technology Law Review (2024) — DOI: 10.36645/mtlr.30.2.tort — Analyzing tort law, product liability, and privacy for Level 3 vehicles as a distinct domain.
• "Liability Rules for Automated Vehicles: Definitions and Details," SMU Science and Technology Law Review (2024) — DOI: 10.25172/smustlr.27.1.5 — Legal fiction of the 'computer driver', excluding human contributory negligence during specific intervals of the handover transition period.
• "Driver Takeover Performance under Different Driving Contexts and Non-Driving Tasks in Level 3 Automated Driving," AHFE International (2026) — DOI: 10.54941/ahfe1007862 — Simulator study of 18 drivers; takeover response is more delayed in urban environments and during non-driving tasks (small-scale empirical study).
• "Research on the Criminal Liability of Manufacturers in Autonomous Vehicle Accidents," Academic Journal of Management and Social Sciences (2026) — DOI: 10.54097/vswsyq46 — Manufacturer criminal liability via product infringement theory, with a legislative proposal to add manufacturers as unitary subjects of traffic offense crimes (China).
• "Negligence Failures and Negligence Fixes," Criminal Law Forum (2023) — DOI: 10.1007/s10609-023-09451-1 — Failures of negligence such as the 'problem of many hands', comparing criminal law models in Singapore, France, and the UK.
• "Regulating driving automation in the European Union – criminal liability on the road ahead?" New Journal of European Criminal Law (2023) — DOI: 10.1177/20322844231213336 — Criticism that the EU pushes technical standards while omitting criminal liability rules, and liability allocation when humans place AI in the driver's seat.
• "Manufacturer Liability for Autonomous Vehicle Accidents in Germany: The Volkswagen Case," Congress Proceedings (2025) — DOI: 10.55843/icl2025cong223p — StVG and ProdHaftG, difficulties of proving causation for opaque AI, proposing a 'software liability document' and compensation fund.
• "Liability Determination for Smart Vehicle Accidents: Based on Black Box and Smart Vehicle Class Assessment," (2025) — DOI: 10.54254/2753-7048/2025.bo23203 — Black boxes are core evidence but have limitations in extreme weather, clarifying liability by L1–L5 grades (the Xu case in China).
• "A Distributed 'Black Box' Audit Trail Design Specification for Connected and Automated Vehicle Data and Software Assurance," SAE Int. J. Transportation Cybersecurity and Privacy (2020) — DOI: 10.4271/11-03-02-0006 — Distributed audit trails for software integrity and provenance assurance (blockchain-based).
• "Measuring data access and re-use in the European Legal Framework for Data ... the case of vehicle data," Open Research Europe (2023) — DOI: 10.12688/openreseurope.16468.1 — Black box and EDR data controlled by insurers and manufacturers, barriers to access rights, and the Data Act.
• "Can there be responsible AI without AI liability?" International Journal of Law and Information Technology (2024) — DOI: 10.1093/ijlit/eaae021 — Discussion on easing the burden of proof (discovery and causation presumption) and moving toward strict liability in the EU AI Liability Directive.
• "A study on the revision of the Compulsory Motor Vehicle Liability Security Act for the era of Autonomous Vehicles," Legal Studies Institute of Chosun University (2024) — DOI: 10.18189/isicu.2024.31.2.149 — The need to overhaul South Korea's Guarantee of Automobile Accident Compensation Act, comparing Germany's doubling of compensation caps and Japan's unmanned driving permit system.
• "State Laws for Autonomous Vehicle Safety, Equity, and Insurance," Journal of Law, Medicine & Ethics (2022) — DOI: 10.1017/jme.2022.96 — Reviewing relevant legislation across 30 U.S. states, noting a lack of impact evaluation data.
• "Who Is at the Wheel? Allocating Tort and Product Liability for Autonomous Vehicle Accidents," (2026) — DOI: 10.2139/ssrn.6543563 — Comparing the UK, US, and Canada, hybrid standards of rebuttable presumption, shifting proof to incident data logs, and insurer-first subrogation.
• "Comparing Tort Liability Frameworks in Autonomous Vehicle Accident Governance," World Electric Vehicle Journal (2026) — DOI: 10.3390/wevj17010032 — Four types of liability frameworks (driver, system, manufacturer/operator, and hybrid) and the institutional dilemmas of each.
• "Not So Fast: A Brief Plea for Muddling Through the Problems of Autonomous Vehicle Liability," Journal of Tort Law (2020) — DOI: 10.1515/jtl-2020-2012 — Attempts to redesign liability frameworks prematurely invite errors; an honest, prudent stance of muddling through step-by-step.