EU AI Liability Directive (proposed, withdrawn)

The AI Liability Directive (AILD) was a legislative proposal by the European Commission, formally the Proposal for a Directive of the European Parliament and of the Council on adapting non-contractual civil liability rules to artificial intelligence, COM(2022) 496, procedure 2022/0303/COD. The proposal document is dated Brussels, 28 September 2022; the EUR-Lex procedure file records its adoption by the Commission on 29 September 2022 under the lead of Justice Commissioner Didier Reynders (Source: eur-lex.europa.eu) (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary). It would have set common rules on two points of national fault-based tort law in claims for damage caused by an AI system: the disclosure of evidence about high-risk AI systems, and the burden of proof (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary). The Commission announced in its 2025 work programme, adopted on 11 February 2025, that it would withdraw the proposal for lack of foreseeable agreement, and the withdrawal was published on 6 October 2025 (Source: iapp.org) (Source: eur-lex.europa.eu).

The proposal was adopted as a package with a revision of the 1985 Product Liability Directive 85/374/EEC (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary); that revision was enacted as Directive (EU) 2024/2853, which treats software, including AI systems, as a product, and which is due to take effect on 9 December 2026 (Source: politico.eu). See AI Liability for the wider set of liability theories and the treatment of the proposal in comparative accounts.

Background

The explanatory memorandum traces the proposal to the Commission's White Paper on AI of 19 February 2020 and its accompanying report on the safety and liability implications of AI, the Council's conclusions of 9 June 2020, and a legislative own-initiative resolution of the European Parliament of 20 October 2020 (2020/2014(INL)) asking the Commission to propose a civil liability regime for AI. Parliament's resolution had suggested a limited strict-liability regime for certain AI-enabled technologies together with an eased burden of proof under fault-based rules (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

The Commission's stated rationale was that the "complexity, autonomy and opacity (the so-called 'black box' effect)" of AI could make it difficult or prohibitively expensive for victims to identify the liable person and prove fault and causation under national rules, that courts adapting those rules case by case would create legal uncertainty for businesses, and that several member states (the memorandum names Czechia, Italy, Malta, Poland and Portugal) were considering national measures that would fragment the internal market. It cited a 2020 Ipsos survey in which liability was the most-cited external obstacle (43%) for companies planning to adopt AI (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

Preparation began in 2018 with the Expert Group on Liability and New Technologies, whose report appeared in November 2019, and drew on comparative-law, behavioural and economic studies. A public consultation from 18 October 2021 to 10 January 2022 received 233 responses: citizens, consumer organisations and academics strongly supported burden-of-proof measures and harmonised strict liability with mandatory insurance, while most business respondents considered strict liability disproportionate and cautioned against a complete shift of the burden of proof. Of three options in the impact assessment (burden-of-proof measures alone; those measures plus strict liability and mandatory insurance for high-risk use cases; or a staged approach deferring strict liability to a review), the Commission chose the staged approach (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

Status and timeline

Sources for the table: the EUR-Lex procedure file (Source: eur-lex.europa.eu), the proposal text (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary) and IAPP's report on the work programme (Source: iapp.org).

The work programme's annex gave the reason as "No foreseeable agreement — the Commission will assess whether another proposal should be tabled or another type of approach should be chosen" (Source: iapp.org).

Scope and definitions

Article 1 limited the directive to non-contractual, fault-based civil claims for damages brought before national courts, for damage caused by an AI system after the end of the transposition period. It did not apply to criminal liability, though the memorandum stated it could apply to state liability. It left national rules on who bears the burden of proof, the standard of proof and the definition of fault untouched except as Articles 3 and 4 provided, and it also left unaffected EU liability rules in transport, rights under national rules implementing the Product Liability Directive, and the Digital Services Act's liability exemptions and due-diligence duties. It was a minimum-harmonisation measure: Article 1(4) allowed member states to adopt or keep rules more favourable to claimants (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary). Recital 10 excluded from harmonisation the definition of fault or causality, types of compensable damage, multiple tortfeasors, contributory conduct, calculation of damages and limitation periods. The explanatory memorandum described rebuttable presumptions as "the least interventionist tool" and stated that the proposal "does not lead to a reversal of the burden of proof" (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

Definitions were tied to the then-pending AI Act: "AI system", "high-risk AI system", "provider" and "user" took their AI Act meanings, and the disclosure and presumption rules for providers and users of high-risk systems cross-referenced the Act's requirements on data quality (Article 10), transparency (Article 13), human oversight (Article 14), and accuracy, robustness and cybersecurity (Article 15), in the numbering of the Commission's 2021 AI Act proposal (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary). A "claim for damages" covered damage caused by an AI system's output or by its failure to produce an output that should have been produced. A "claimant" could be the injured person, a person subrogated to their rights (such as an insurer or heir), or a person acting on behalf of injured persons. Recital 15 excluded cases where an AI system only provided information or advice that a human then acted on (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

Key provisions

Disclosure of evidence (Article 3). National courts were to be empowered to order a provider or user of a high-risk AI system suspected of having caused damage to disclose relevant evidence, on the request of a claimant or a potential claimant who had first asked for it and been refused. The requester had to present facts and evidence sufficient to support the plausibility of a claim, and a court could order disclosure from a non-defendant only where the claimant had made all proportionate attempts to obtain the evidence from the defendant. Courts could also order preservation of evidence, and had to limit both to what was necessary and proportionate. Trade secrets were to be protected under Directive (EU) 2016/943, and confidential information such as that related to public or national security was also to be weighed. Under Article 3(5), a defendant's failure to comply with a disclosure or preservation order led the court to presume non-compliance with the relevant duty of care; the presumption was rebuttable. A refusal of the pre-litigation request did not itself trigger the presumption (recital 17) (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

Presumption of causality (Article 4). Courts were to presume the causal link between a defendant's fault and the output of an AI system, or its failure to produce an output, where three conditions were all met: (a) the claimant demonstrated, or the court presumed under Article 3(5), fault consisting of non-compliance with a duty of care in Union or national law directly intended to protect against the damage that occurred; (b) it was reasonably likely, on the circumstances, that the fault influenced the output or failure to produce one; and (c) the claimant demonstrated that the output or failure gave rise to the damage (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary). Recital 25 illustrated condition (b): breaching a duty to limit a system's perimeter of operation, where damage occurred outside that perimeter, would qualify; failing to file documents or register with an authority would not (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

The presumption was qualified by type of system and defendant:

• Against a provider of a high-risk system, fault under condition (a) could be shown only by non-compliance with listed AI Act requirements — training-data quality, transparency, human oversight, accuracy/robustness/cybersecurity, or failure to take corrective action or withdraw the system — taking into account the provider's risk-management system (Article 4(2)).
• Against a user of a high-risk system, fault was shown by failing to use or monitor the system according to its instructions, or by exposing it to irrelevant input data under the user's control (Article 4(3)).
• For a high-risk system, the presumption did not apply where the defendant showed that sufficient evidence and expertise were reasonably accessible for the claimant to prove causation (Article 4(4)).
• For a system that was not high-risk, the presumption applied only where the court considered it excessively difficult for the claimant to prove causation (Article 4(5)).
• Against a defendant using the system in a personal, non-professional activity, it applied only where the defendant materially interfered with the system's operating conditions or was required and able to set them and failed to do so (Article 4(6)).

The defendant could rebut the presumption in every case (Article 4(7)) (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

Review (Article 5). Five years after transposition the Commission was to report on the effects of Articles 3 and 4 and, in particular, evaluate "the appropriateness of no-fault liability rules for claims against the operators of certain AI systems" and the need for insurance coverage, supported by a monitoring programme to which member states would report annually. The memorandum explained the deferral on the ground that AI-equipped products and services that could put life, health and property at risk were "not yet widely available on the market" (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

Other provisions (Articles 6–9). Article 6 added the directive to Annex I of the Representative Actions Directive (EU) 2020/1828, opening it to collective consumer actions. Article 7 set a two-year transposition period, and Article 8 entry into force on the twentieth day after publication in the Official Journal (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

Estimated effects

The Commission's impact assessment, relying on a Deloitte economic study, estimated "as a conservative estimate" that targeted harmonisation would raise the production value of relevant cross-border trade by 5 to 7% and generate additional EU AI market value of roughly EUR 500 million to EUR 1.1 billion in 2025. The proposal had no EU budget implications (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary).

Relation to the Product Liability Directive

The memorandum described the two proposals as a package dealing with different types of liability: the Product Liability Directive covers producers' no-fault liability for defective products, "leading to compensation for certain types of damages, mainly suffered by individuals", while the AI Liability Directive covered national claims "mainly based on the fault of any person" for any type of damage and any type of victim (Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary). Only the product-liability revision was enacted (Directive (EU) 2024/2853) (Source: politico.eu).

Reactions

On the withdrawal, Member of the European Parliament Axel Voss told IAPP that the directive was an ex post mechanism distinct from the AI Act's preventive aim, and attributed the reversal to "pressure from industry lobbyists." Peter Craddock of Keller and Heckman said forgoing debate was preferable to repealing legislation after passage, while noting the unresolved question of how harmonised liability would now be achieved. CCIA Europe welcomed the shelving (Source: iapp.org).

Calls for a successor (2026)

In a note to the Commission dated 16 September 2026, four members of the European Parliament — Kim van Sparrentak (Greens), Axel Voss (EPP), Brando Benifei (S&D) and Michael McNamara (independent) — asked the Commission to introduce an "AI Liability Act" to "fill the legislative gaps currently existing for these kinds of incidents, so no people or European businesses are left empty-handed against these big providers when they cause harm," naming OpenAI and Anthropic as examples of providers that could be held liable when their technology is deployed in unforeseen ways. POLITICO Europe, which saw the note, reported support from members of the EPP, S&D and Renew groups and linked the call to the 2026 agent intrusion incidents, including the Hugging Face intrusion by OpenAI-powered agents (see AI Agent Containment Incidents, 2026). Mario Mariniello of Bruegel said the existing arrangement "basically grants no rights to people who are harmed by artificial intelligence" (Source: politico.eu).

Relationships

• supports: Proposal for an AI Liability Directive, COM(2022) 496 — Source Summary — primary text of the proposal (explanatory memorandum, recitals, Articles 1–9)
• depends-on: EU AI Act (Regulation 2024/1689) — the disclosure and presumption rules for high-risk systems were keyed to AI Act requirements
• instance-of: AI Liability — EU fault-based harmonisation approach, withdrawn
• related: AI and Tort Liability — tort doctrine applied to AI
• related: Digital Services Act (EU) — liability exemptions left unaffected by Article 1(3)(c)
• related: EU Cyber Resilience Act — named in the explanatory memorandum as complementary; grouped with the liability directives among surrounding EU instruments in Regulating Under Uncertainty: Governance Options for Generative AI (G'sell, 2024)
• related: AI LEAD Act (S. 2937) — US products-liability framework compared with the EU arrangement in Inside My AI Law & Policy Class 11: When AI Fails, Who Pays? (Farahany, September 2025)
• related: European Parliament — 2020 resolution requesting a civil liability regime; source of the 2026 call for a successor instrument
• related: AI Agent Containment Incidents, 2026 — incidents cited in the 2026 call