Should manufacturers and operators of autonomous systems be held legally liable for unintended harms caused by their automation and robotics deployments?
6 peer-reviewed studies for, 4 against
Key takeaways
On balance, the evidence supports holding manufacturers and operators of autonomous systems legally liable for unintended harms.
Legal scholarship argues that existing product-liability and tort doctrines can already hold makers responsible when no human directs the machine.
The EU's revised Product Liability Directive (2024/2853) treats software and AI as products subject to strict, no-fault producer liability.
Critics warn that assigning fault to self-learning systems is hard and that aggressive new liability rules could delay safety-improving automation.
None of the opposing sources argue for no liability. The open question is how to allocate it, such as shared regimes or no-fault insurance.
Stronger evidence: For
Verdict
On balance the evidence supports holding manufacturers and operators of autonomous systems legally liable for unintended harms, though the design of that liability regime matters enormously. The "for" case is anchored by both scholarship and enacted law. Vladeck (Washington Law Review, 2014) argues that traditional product-liability doctrine — a product that fails to perform as safely as an ordinary consumer would expect — already supplies a workable basis for holding the makers of autonomous systems responsible even when no human is directing the machine. The European Union has since gone further in law: Directive (EU) 2024/2853, the revised Product Liability Directive, explicitly extends strict, no-fault producer liability to software and AI systems, and keeps producers on the hook for defective components even after deployment where they control updates. Custers, Lahmann and Scott (AI & Society, 2025) supply the animating rationale — real "liability gaps" leave victims of AI harm under-compensated, and because AI is built and run through overlapping processes with many partial-control actors, responsibility should be shared rather than allowed to fall through the cracks. The "against" evidence, however, is a serious and legitimate check on how far to go. Villasenor (Brookings, 2014) warns that existing products-liability law is already adaptive and that front-loading aggressive new liability rules should not be a precondition to rollout, lest it delay automation that saves lives. That safety dividend is real: NHTSA (DOT HS 812 115, 2015) found the driver was the critical reason in 94% of crashes, so removing human error can reduce overall harm — an argument that blanket manufacturer blame can be misplaced. And both Kubica (American Journal of Comparative Law, 2022) and Long, Zhao and Cai (World Electric Vehicle Journal, 2026) show that assigning fault to self-learning systems is genuinely hard, and that alternative or composite regimes — no-fault insurance, shared enterprise responsibility — may outperform a single strict-manufacturer rule. Critically, none of the "against" sources argue for NO liability; they argue for carefully engineered liability. The synthesis: manufacturers and operators should bear liability for the harms their automation causes, because someone must answer to injured people and diffuse responsibility otherwise leaves them uncompensated. But the strongest regime is not naive strict liability on the maker alone — it is a shared/composite framework (potentially backed by no-fault insurance) that preserves accountability without freezing beneficial deployment. The claim that they "should be held legally liable" is supported; the open question is the exact allocation, not whether liability attaches.
Herbosch provides a comprehensive law-and-economics analysis showing that AI agents do not require fundamental overhaul of tort law but targeted refinements to existing liability frameworks. The paper argues that manufacturers and operators can be held liable through traditional product and tort liability doctrines when AI systems cause harm.
Kubica's comparative survey shows traditional product liability struggles with autonomous-vehicle software defects and causation, and canvasses alternatives to manufacturer liability — notably a no-fault insurance model (akin to workers' compensation) triggered once autonomous vehicles reach around 25% market penetration, plus common-enterprise liability and two-tier insurance caps. It supports the view that well-designed alternative regimes can outperform strict manufacturer liability.
Herbosch provides a comprehensive law-and-economics analysis showing that AI agents do not require fundamental overhaul of tort law but targeted refinements to existing liability frameworks. The paper argues that manufacturers and operators can be held liable through traditional product and tort liability doctrines when AI systems cause harm.
David Vladeck argues that traditional tort and product-liability principles can and should extend to autonomous AI systems, apportioning responsibility among designers, programmers, and manufacturers. When a machine fails to perform as safely as an ordinary consumer would expect, product-liability doctrine supplies a workable basis for holding its makers responsible even when no human "principal" was directing the machine.
The European Commission proposed easing victims' burden of proof with a rebuttable "presumption of causality" linking an AI system's fault to the harm it caused, plus court-ordered disclosure of evidence about high-risk AI systems. It reflects an official push to make manufacturers and operators answerable for AI harms. (The Commission later withdrew the proposal in 2025 after no agreement was reached, so it stands as a proposal, not enacted law.)
The revised EU Product Liability Directive — now enacted law repealing the 1985 directive — explicitly treats software and AI systems as "products" subject to no-fault (strict) producer liability, including cloud- and SaaS-delivered software. Producers can remain liable even after a product leaves their control where they retain authority over software updates, keeping manufacturers accountable for defective digital and AI components.
Custers, Lahmann and Scott document multiple "liability gaps" where existing rules leave victims of AI harm under-compensated. Because AI is developed, deployed, and used through overlapping, co-created processes in which many actors hold partial control, they argue those actors should bear shared responsibilities and fiduciary duties — turning liability gaps into overlapping liability rather than leaving injured people without recourse.
Provides quantified biomechanical analysis of robot accident injury thresholds using finite element modeling. Demonstrates that injury severity can be measured and predicted from robot parameters, establishing scientific basis for accountability when manufacturers fail to prevent foreseeable harms from autonomous robotic deployments.
Kubica's comparative survey shows traditional product liability struggles with autonomous-vehicle software defects and causation, and canvasses alternatives to manufacturer liability — notably a no-fault insurance model (akin to workers' compensation) triggered once autonomous vehicles reach around 25% market penetration, plus common-enterprise liability and two-tier insurance caps. It supports the view that well-designed alternative regimes can outperform strict manufacturer liability.
This comparative analysis identifies four liability regimes — driver, system, manufacturer/operator, and composite — and the institutional problems each creates, including "blurred boundaries of liability" and weak fact-finding mechanisms under manufacturer- or operator-focused rules. The authors argue that pinning fault on autonomous systems is genuinely hard and that a single manufacturer-liability rule is problematic, favoring composite approaches over any one actor bearing all the risk.
John Villasenor argues that existing products-liability law is already "remarkably adaptive" and that no expansive new manufacturer-liability regime is needed for automation. A guiding principle he offers is that "preemptively resolving liability issues should not be a precondition to commercial rollout" — front-loading aggressive liability rules would needlessly delay the deployment of safety-improving automation.
In this federal survey, NHTSA assigned the critical reason — the last event in the crash causal chain — to the driver in 94% of crashes (plus or minus 2.2%), versus about 2% to the vehicle and 2% to the environment. Because most crashes trace to human error, automation that removes the human driver could reduce overall harm, weakening the case for broad manufacturer liability. NHTSA cautions this is the last event in the chain, not an assignment of blame or a complete cause.
The debate
Lisa: When an autonomous system injures someone, a person should be able to recover, and the maker is the obvious place responsibility lands. Vladeck made this case back in 2014 in the Washington Law Review: product-liability law already asks whether a product performed as safely as an ordinary consumer would expect. You do not need a human at the wheel for that test to work. And the EU has now written it into law — the 2024 Product Liability Directive treats AI and software as products under strict liability.
James: I do not disagree that victims should recover. My worry is what strict manufacturer liability does to deployment. Villasenor at Brookings argued existing products law is already remarkably adaptive, and that resolving every liability question up front should not be a precondition to rollout. If we over-index on punishing makers, we delay automation that is safer than the status quo.
Lisa: But "safer than the status quo" cuts the other way too. If the technology really is safer, manufacturers should be happy to stand behind it. The reason liability matters is the gap Custers and colleagues documented in 2025 — real cases where nobody is clearly responsible and the injured party just eats the loss. That is the outcome we should refuse to accept.
James: Agreed that gaps are the enemy. I just think the fix is not always "blame the manufacturer." NHTSA found the driver was the critical reason in 94% of crashes. When you remove the human, you remove most of the harm — so a rule that automatically routes every remaining accident to the maker is assigning fault where the causal story is genuinely murky.
Lisa: The 94% figure is about today's human-driven fleet, though, and NHTSA itself says it is the last event in the chain, not blame. Once a machine is making the decisions, the design choices behind those decisions are exactly what product liability is meant to police.
James: Fair. But look at Kubica's comparative work and the 2026 World Electric Vehicle Journal analysis — both show that pinning fault on a self-learning system is hard, and that composite or no-fault insurance models can compensate victims faster without a courtroom fight over an opaque algorithm. Workers' compensation is the analogy: you get paid without proving the employer was negligent.
Lisa: I actually like no-fault insurance — as a complement, not a replacement. Someone still has to fund the pool and bear the incentive to build safer systems, and that someone is the manufacturer and operator. Strip out their liability entirely and you strip out the reason to invest in safety.
James: Then I think we are closer than it sounds. Liability should attach — I just want it engineered as shared responsibility across the actors who actually controlled the risk, backed by insurance, rather than a blunt strict-liability hammer on the maker alone.
Lisa: That I can agree with. Hold manufacturers and operators liable, yes — but through a composite regime that compensates victims reliably and keeps the incentive to make automation safer. The question was never whether they answer for the harm; it is how.
Frequently asked questions
Who is responsible when an autonomous system or robot causes harm?
Sources supporting liability argue that designers, programmers, manufacturers and operators can be held responsible through traditional product and tort liability. One analysis argues that because many actors hold partial control over AI, they should share responsibility so injured people are not left without recourse.
Does EU law make companies liable for harm caused by AI?
Yes. Directive (EU) 2024/2853 explicitly treats software and AI systems as products subject to no-fault producer liability, including cloud-delivered software. Producers can remain liable after release when they control software updates. A separate proposed AI Liability Directive was withdrawn in 2025 and was never enacted.
Could strict manufacturer liability slow down self-driving cars?
Some analysts warn that it could. A Brookings paper argues that resolving liability issues in advance should not be a precondition to rollout, because heavy liability rules could delay automation that improves safety.
What alternatives to manufacturer liability have been proposed for autonomous vehicles?
Comparative legal studies describe several alternatives. These include a no-fault insurance model similar to workers' compensation, common-enterprise liability, two-tier insurance caps, and composite regimes that spread responsibility among drivers, systems, manufacturers and operators rather than placing it on one actor.