AI Liability and Product Safety Law

AI liability and product safety law addresses one of the most challenging legal questions of the emerging technology era: who is responsible when an AI system causes harm, and how can that responsibility be proven when AI systems are opaque, complex, and involve multiple actors across a long supply chain? The framework is evolving across jurisdictions. In the EU, three instruments form a layered approach: the AI Act (2024) classifies AI systems by risk and imposes obligations on high-risk providers; the AI Liability Directive (AILD, proposed 2022) introduces a presumption of causation and disclosure of evidence to help claimants prove AI-related harm; and the revised Product Liability Directive (PLD, 2024) explicitly extends strict liability to software and AI systems. In the US, traditional tort law (negligence) and product liability (the Restatement Third of Torts: Products Liability) apply, though AI's opacity and multi-actor nature challenge these frameworks. The UK's Consumer Protection Act 1987 provides strict product liability.

The framework confronts two structural challenges. The first is allocation: AI systems involve developers (who design and train models), deployers (who integrate and use them), operators (who manage deployment), and users (who provide inputs) — and harm may result from any combination of their actions. The second is proof: AI systems are often opaque (deep neural networks are not directly interpretable), complex (involving training data, model architecture, fine-tuning, and deployment context), and causally distant from the harm (an output may depend on training data collected years earlier by a different organization).

This page covers the liability frameworks, product liability for AI, AI-specific rules, risk allocation across the AI supply chain, and practical compliance for AI providers and deployers.

Traditional tort liability — negligence — requires the claimant to prove duty, breach, causation, and damage. For AI systems, each element presents challenges. Duty: who owed a duty to the harmed party — the developer, the deployer, or the user? Breach: what standard of care applies to AI development and deployment, and was it met? Causation: was the harm caused by the AI system's output, the training data, the deployment context, or the user's input? Damage: what is the nature and extent of the harm? These challenges make traditional negligence claims difficult to pursue for AI-related harm.

Product liability — strict liability for defective products — offers an alternative path because it does not require proof of negligence. The claimant must prove the product was defective, the defect caused the harm, and the product was in the stream of commerce. The question for AI is whether an AI system is a "product" at all (software has historically been treated as a service in some jurisdictions), and if so, what constitutes a "defect" — a flaw in the training data, the model architecture, the output, or the post-deployment updates?

AI-specific liability rules are the newest layer. The EU's AILD introduces a presumption of causation: once a claimant shows harm, an AI system was involved, and a likely fault (a breach of a duty of care relevant to the harm), the court presumes the causal link between the fault and the harm, shifting the burden to the defendant to rebut. The AILD also provides for disclosure of evidence: courts can order providers and deployers to disclose relevant information about the AI system's development and operation, subject to trade secret protections.

The revised EU Product Liability Directive (2024) represents a significant expansion of the 1985 PLD. It explicitly brings software and AI systems within the scope of "products" subject to strict liability, addressing the prior ambiguity about whether software was a product or a service. The revision also extends the limitation period (the time during which claims can be brought) to 10 years, introduces a discoverability rule (the limitation period runs from when the claimant discovered or should have discovered the defect, damage, and causal link), and removes the previous €500 threshold for property damage claims.

The defect standard remains the "legitimate expectations" test: a product is defective if it does not provide the safety that a person is entitled to expect, taking into account the product's presentation, the use to which it could reasonably be put, and the time the product was put into circulation. For AI systems, this raises novel questions: is an AI system defective if its training data was biased, if the model architecture produces errors, if the output is correct but unexpected, or if the system was not updated after a post-deployment risk emerged? The PLD's reference to "the time the product was put into circulation" may be difficult to apply to AI systems that learn and update over time.

In the United States, the Restatement (Third) of Torts: Products Liability distinguishes manufacturing defects (the product deviates from its intended design), design defects (the product's design is unreasonably dangerous), and warning defects (inadequate instructions or warnings). For AI, design defects are the most likely claim: a model that produces discriminatory outputs, hallucinates confidently, or fails in foreseeable use conditions may be alleged to have an unreasonably dangerous design. Whether US courts will treat AI as a "product" for strict liability purposes remains an open question; current case law is sparse but developing.

The EU AI Liability Directive (AILD) is designed to address the proof difficulties that AI opacity and complexity create. The core mechanism is a presumption of causation: if a claimant can demonstrate (a) that the defendant breached a duty of care relevant to the harm (such as a GDPR obligation or an AI Act requirement), (b) that the AI system was capable of causing the harm, and (c) that the harm occurred, then the court presumes that the defendant's fault caused the harm. The defendant can rebut this presumption by showing the breach did not cause the harm — but the burden shifts to the defendant, which significantly lowers the barrier for claimants.

The AILD also provides for disclosure of evidence. Courts can order AI providers, deployers, and other relevant parties to disclose information necessary to prove or defend an AI-related claim, including information about the AI system's design, training data, testing, and operation. Trade secrets are protected through proportionality assessments and confidentiality orders, but the disclosure mechanism addresses a critical asymmetry: AI providers hold the information needed to prove liability, and claimants cannot access it without compulsory disclosure.

The EU AI Act (2024) contributes to liability indirectly by establishing duties of care. High-risk AI providers must implement risk management systems, ensure data governance, provide transparency, maintain human oversight, ensure accuracy and robustness, keep logs, and conduct post-market monitoring. Non-compliance with these obligations can constitute the "fault" that triggers the AILD's presumption of causation — linking the regulatory framework to the liability framework. The AI Act also imposes substantial administrative fines for non-compliance, up to €35 million or 7% of worldwide annual turnover for the most serious violations.

The AI supply chain involves multiple actors whose contributions to harm can be difficult to disentangle. Developers design the model architecture and curate training data. Fine-tuners adapt base models to specific tasks or domains. Deployers integrate the model into an application or service and configure it for use. Operators manage the deployment infrastructure and monitor performance. Users provide inputs and interact with the system's outputs. Harm may result from any combination: biased training data, a flawed architecture, an inappropriate fine-tuning dataset, a misconfigured deployment, inadequate monitoring, or a user's misuse.

Risk allocation among these actors is typically addressed through contracts (indemnification clauses, limitation of liability, warranties), insurance (professional liability, cyber liability, and emerging AI-specific coverage), and law (strict liability, negligence, and the new AI-specific rules). The challenge is that contractual allocation may not bind third-party claimants (a user harmed by the AI can sue any actor in the chain, regardless of their internal agreements), and insurance coverage for AI-related harm is still maturing — standard policies may exclude AI-related risks, and AI-specific coverage is not yet widely available or standardized.

Mandatory insurance is an emerging consideration. The EU AI Act's high-risk classification has prompted discussion of whether mandatory liability insurance should be required for high-risk AI deployments, similar to mandatory insurance in aviation, automotive, or medical malpractice. This would ensure that harmed parties can recover even when individual providers are insolvent or judgment-proof, but it would also add cost and complexity to AI deployment.

AI providers and deployers should adopt a structured compliance approach. First, assess the AI system's risk classification: determine whether it falls under the AI Act's high-risk categories (Annex III covers biometric identification, critical infrastructure, education and employment, essential services, law enforcement, migration and asylum, and justice and democratic processes), whether it qualifies as a product under the PLD, and what liability exposures it presents (product liability, negligence, AI-specific). The classification determines which obligations apply and which liability frameworks govern.

Second, comply with applicable obligations: implement the AI Act's high-risk requirements (risk management, data governance, transparency, human oversight, accuracy, logging, post-market monitoring), ensure the product meets PLD safety expectations, and prepare for AILD evidence disclosure by maintaining documentation that can be produced in litigation.

Third, document everything: maintain records of risk assessments, data governance decisions, testing and validation, model cards, logging, transparency disclosures, human oversight mechanisms, and post-market monitoring. This documentation is the primary defense in liability litigation — it demonstrates that the provider met its duty of care and allows the provider to rebut a presumption of causation if one arises. The AILD's disclosure mechanism means this documentation may be ordered produced in court, so it should be comprehensive and accurate.

Fourth, allocate and insure risk: use contracts to allocate responsibility among developers, deployers, and operators (indemnification, limitation of liability, warranties); obtain appropriate insurance coverage (professional liability, cyber liability, and AI-specific coverage as it emerges); and monitor post-market for serious incidents, model drift, and feedback that may indicate emerging risks. The AI Act requires serious incident reporting to authorities; prompt reporting and remediation demonstrate diligence and may mitigate liability.