The AI Liability Directive proposal is a draft Directive of the European Parliament and of the Council "on adapting non-contractual civil liability rules to artificial intelligence", issued by the European Commission as COM(2022) 496 final. It would have harmonised two narrow aspects 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 proving the causal link between a defendant's fault and the AI system's output. It left fault, damage and most procedural rules to national law, and deferred the question of no-fault (strict) liability and mandatory insurance to a review five years after transposition. The Commission withdrew the proposal in 2025. The structured legislation page, including the procedural history and withdrawal, is EU AI Liability Directive (proposed, withdrawn).
| Field | Detail |
|---|---|
| Document | Proposal for a Directive of the European Parliament and of the Council on adapting non-contractual civil liability rules to artificial intelligence (AI Liability Directive) |
| Identifiers | COM(2022) 496 final; procedure 2022/0303 (COD); CELEX 52022PC0496; accompanied by SEC(2022) 344 final and SWD(2022) 318, 319 and 320 final |
| Author | European Commission |
| Date | Document dated Brussels, 28 September 2022; the EUR-Lex procedure file records adoption by the Commission on 29 September 2022, with Didier Reynders as leading person (Source: eur-lex.europa.eu) |
| Legal basis | Article 114 of the Treaty on the Functioning of the European Union (internal market) |
| Instrument | Directive, minimum harmonisation; ordinary legislative procedure |
| Contents | Explanatory memorandum (five parts); 33 recitals; 9 articles; 40 footnotes |
| Outcome | Withdrawn (withdrawal published 6 October 2025, CELEX 52025XC05423) |
Provenance
The raw file is the English HTML text of the proposal on EUR-Lex, the EU's official legal database, captured on 2026-09-24. Authenticity was checked against the canonical host (europa.eu domain), the document number and procedure number printed on the letterhead (COM(2022) 496 final; 2022/0303(COD)), the EUR-Lex page metadata (document ID 52022PC0496) and the separate EUR-Lex procedure file. Site navigation was removed; the text runs from the Commission letterhead through the last footnote, including some extraction damage (dropped spaces between words, footnotes run together on one line). Bracketed placeholders in the original, such as "[the AI Act]", "[two years after the entry into force]" and "[DATE five years after the end of the transposition period]", are part of the Commission text. The proposal cross-refers to articles of the Commission's 2021 AI Act proposal, COM(2021) 206 (footnote 31), whose numbering differs from the enacted AI Act. The document contains no personal data of private individuals.
Problem statement (explanatory memorandum, part 1)
The memorandum opens with a 2020 Ipsos enterprise survey in which liability ranked among the top three barriers to AI use by European companies and was the most-cited external obstacle (43%) for companies planning but not yet using AI. It traces the proposal to President Ursula von der Leyen's political guidelines, the White Paper on AI of 19 February 2020 and the accompanying report on safety and liability (COM(2020) 64), the Council's conclusions of 9 June 2020, and the European Parliament's own-initiative resolution of 20 October 2020 (2020/2014(INL)) requesting a civil liability regime for AI under Article 114 TFEU, as well as Parliament's resolution of 3 May 2022 on AI in a digital age.
The Commission's diagnosis is that national fault-based rules require a victim to prove a wrongful act or omission by the person who caused the damage, and that "the specific characteristics of AI, including complexity, autonomy and opacity (the so-called 'black box' effect)" may make identifying the liable person and proving the claim difficult or prohibitively expensive. It predicts that courts would adapt existing rules ad hoc, producing legal uncertainty that businesses, particularly SMEs and cross-border traders, could not price or insure, and that member states acting alone would fragment the internal market. The memorandum names national AI strategies in Czechia, Italy, Malta, Poland and Portugal as mentioning initiatives to clarify liability.
The stated objective is "to promote the rollout of trustworthy AI" by giving victims of AI-caused damage protection equivalent to victims of damage caused by products in general, reducing businesses' legal uncertainty, and preventing fragmented national adaptations.
Place in the Commission's AI package
The memorandum presents the proposal as one of three work streams, alongside the AI Act proposal and a revision of sectoral and horizontal product-safety rules. It describes safety and liability as "two sides of the same coin": AI Act requirements reduce risk without eliminating it, and liability rules apply where the risk materialises, while also giving an economic incentive to comply. Non-compliance with AI Act requirements for high-risk systems is described as "an important element triggering the alleviations of the burden of proof".
The proposal was adopted together with a proposal to revise the Product Liability Directive 85/374/EEC. The memorandum divides the ground between them: the Product Liability Directive covers producers' no-fault liability for defective products and certain types of damage suffered mainly by individuals, while this proposal covers national claims "mainly based on the fault of any person" for any type of damage and any type of victim. The memorandum also names the Cyber Resilience Act proposal (COM(2022) 454) as complementary, and states that the proposal does not affect the Digital Services Act's liability exemptions and due-diligence rules for intermediaries.
Legal basis, subsidiarity and proportionality (part 2)
The legal basis is Article 114 TFEU. Without EU rules, the memorandum says, providers, operators, users and injured persons "would be faced with 27 different liability regimes".
The proposal is built as a staged approach. The first stage is limited to burden-of-proof measures. Of the tools national laws use to ease the burden of proof (strict liability, reversal of the burden, irrebuttable or rebuttable presumptions, per footnote 23), the Commission chose "rebuttable presumptions as the least interventionist tool", and states that the proposal "does not lead to a reversal of the burden of proof, to avoid exposing providers, operators and users of AI systems to higher liability risks, which may hamper innovation". The second stage is a review of whether to harmonise other elements, "including for situations where strict liability would be more appropriate, as requested by the European Parliament", possibly coupled with mandatory insurance.
A directive was chosen over a recommendation on the reasoning that non-binding instruments have unpredictable take-up, particularly in private law, which the memorandum characterises as marked by long-standing national traditions.
Consultation, expertise and impact assessment (part 3)
- Public consultation. A dedicated online consultation ran from 18 October 2021 to 10 January 2022 and received 233 responses from 21 member states and third countries. Citizens, consumer organisations and academic institutions strongly supported burden-of-proof measures and harmonised strict liability with mandatory insurance. Businesses were divided; the majority of business respondents considered strict liability disproportionate, easing the burden of proof drew more support (particularly among SMEs), and businesses cautioned against a complete shift of the burden of proof.
- Expertise. Preparatory work began in 2018 with the Expert Group on Liability and New Technologies (New Technologies Formation), whose November 2019 report assessed AI's challenges to national liability rules. Three external studies followed: a comparative-law study (Karner, Koch and Geistfeld, 2021), a behavioural study (Kantar, 2021) and an economic study (Deloitte, 2021).
- Impact assessment. The Regulatory Scrutiny Board gave a positive opinion with comments at its meeting of 6 April 2022. Three options were assessed: (1) three burden-of-proof measures; (2) option 1 plus harmonised strict liability for AI use cases with a particular risk profile, coupled with mandatory insurance; (3) a staged approach of option 1 followed by a review of strict liability and insurance. Option 3 ranked highest in the multi-criteria analysis and was chosen.
- Estimated effects. The economic study concluded, "as a conservative estimate", that targeted harmonisation would raise the production value of relevant cross-border trade by 5 to 7% against the baseline, and the Commission estimated additional EU-27 AI market value of roughly EUR 500 million to EUR 1.1 billion in 2025.
- Fundamental rights. The memorandum links the proposal to the right to an effective remedy and fair trial (Article 47 of the Charter) and to rights to life, physical and mental integrity and property, and states that it "does not create or harmonise the duties of care or the liability" of entities regulated by the AI Act, GDPR, DSA or non-discrimination law.
The proposal has no implications for the EU budget (part 4). Part 5 sets out a monitoring plan drawing on member-state reporting, AI Act market-surveillance information (Article 62 of the AI Act proposal), AI Act evaluation reports, product-safety analyses and the Commission's report under the Motor Insurance Directive on autonomous and semi-autonomous vehicles.
Scope and definitions (Articles 1–2)
Article 1(1) lays down common rules on (a) the disclosure of evidence on high-risk AI systems to substantiate a non-contractual fault-based civil claim for damages, and (b) the burden of proof in such claims brought before national courts for damage caused by an AI system. It applies only to damage occurring after the end of the transposition period and not to criminal liability; the memorandum states it "may be applicable with respect to state liability".
Article 1(3) leaves unaffected: EU rules on liability in transport; rights under national rules implementing the Product Liability Directive 85/374/EEC; the Digital Services Act's liability exemptions and due-diligence obligations; and national rules on which party bears the burden of proof, the standard of proof and the definition of fault, except as Articles 3 and 4 provide. Article 1(4) allows member states to adopt or keep rules more favourable to claimants. Recital 14 gives national reversals of the burden of proof and national strict-liability regimes as examples.
Recital 10 lists the general aspects of civil liability the directive does not harmonise: the definition of fault or causality, types of compensable damage, distribution of liability among multiple tortfeasors, contributory conduct, calculation of damages and limitation periods. Recital 15 excludes damage caused by a human assessment and act where the AI system only provided information or advice, because causality there is no harder to prove than without AI.
Article 2 takes the definitions of "AI system", "high-risk AI system", "provider" and "user" from the AI Act. A "claim for damages" is a non-contractual fault-based claim for damage caused by an AI system's output or its failure to produce an output "where such an output should have been produced". A "claimant" includes the injured person, a person who has succeeded or been subrogated to the injured person's rights (the memorandum names insurers and heirs), and a person acting on behalf of one or more injured persons. A "duty of care" is "a required standard of conduct, set by national or Union law, in order to avoid damage to legal interests recognised at national or Union law level, including life, physical integrity, property and the protection of fundamental rights."
Disclosure of evidence (Article 3)
National courts are to be empowered to order a provider, a person subject to a provider's obligations under Article 24 or 28(1) of the AI Act, or a user to disclose relevant evidence about a specific high-risk AI system suspected of having caused damage. A potential claimant must first have asked the person for the evidence and been refused, and must present "facts and evidence sufficient to support the plausibility of a claim for damages". Disclosure from a provider or user who is not the defendant may be ordered only where the claimant has made "all proportionate attempts" to obtain the evidence from the defendant (Article 3(2)). Courts may also order preservation of evidence (Article 3(3)).
Disclosure and preservation are limited to what is "necessary and proportionate". Courts must weigh the legitimate interests of all parties, including third parties, in particular trade secrets under Directive (EU) 2016/943 and confidential information such as information related to public or national security, may take measures to preserve the confidentiality of trade secrets used in proceedings, and must give the person ordered to disclose appropriate procedural remedies (Article 3(4)). Recital 17 states that the refusal of a pre-litigation request does not itself trigger the presumption of non-compliance.
Under Article 3(5), where a defendant fails to comply with a court order to disclose or preserve evidence, the court "shall presume the defendant's non-compliance with a relevant duty of care" that the evidence was intended to prove. The defendant may rebut the presumption. The memorandum describes the rule as a procedural tool meant "to promote disclosure but also to expedite court proceedings".
Presumption of causality (Article 4)
Article 4(1) requires national courts to presume "the causal link between the fault of the defendant and the output produced by the AI system or the failure of the AI system to produce an output" where all three conditions are met:
- (a) the claimant has demonstrated, or the court has presumed under Article 3(5), the fault of the defendant (or of a person for whose behaviour the defendant is responsible), consisting in non-compliance with a duty of care in Union or national law "directly intended to protect against the damage that occurred";
- (b) "it can be considered reasonably likely, based on the circumstances of the case, that the fault has influenced the output produced by the AI system or the failure of the AI system to produce an output";
- (c) the claimant has demonstrated that the output or failure to produce an output gave rise to the damage.
The presumption thus covers only the link between fault and output. The claimant still proves fault (or obtains the Article 3(5) presumption) and the link between output and damage. Recitals 22 and 25 give examples of the "directly intended to protect" and "reasonably likely" tests: breaching instructions of use meant to prevent physical harm can support the presumption in a personal-injury claim, and a breach of a duty to limit a system's perimeter of operation is reasonably likely to have influenced an output causing damage outside that perimeter, whereas a failure to file documents or register with an authority would not qualify.
The presumption is then qualified by type of defendant and system:
| Provision | Case | Rule |
|---|---|---|
| Art. 4(2) | Claim against a provider of a high-risk AI system (or person with provider obligations) | Fault under condition (a) is met only by showing non-compliance with listed AI Act requirements, taking into account the risk-management system: training, validation and testing data quality (Art. 10(2)–(4)); transparency (Art. 13); effective human oversight (Art. 14); accuracy, robustness and cybersecurity (Arts. 15 and 16(a)); or failure to take corrective action or withdraw or recall the system (Arts. 16(g) and 21) |
| Art. 4(3) | Claim against a user of a high-risk AI system | Fault is met where the user did not use or monitor the system according to its instructions of use, or suspend or interrupt its use (Art. 29), or exposed it to input data under its control not relevant to the system's intended purpose (Art. 29(3)) |
| Art. 4(4) | High-risk system | Presumption does not apply where the defendant shows that sufficient evidence and expertise is reasonably accessible for the claimant to prove causation |
| Art. 4(5) | System that is not high-risk | Presumption applies only where the court considers it excessively difficult for the claimant to prove causation |
| Art. 4(6) | Defendant using the system in a personal, non-professional activity | Presumption applies only where the defendant materially interfered with the system's conditions of operation, or was required and able to determine them and failed to do so |
| Art. 4(7) | All cases | Defendant may rebut the presumption |
Article references in the table are to the AI Act as proposed in 2021. Recital 28 adds that a claimant relying on Article 4(5) need not explain the characteristics of the AI system or how they make causation harder to establish. Recital 29 explains Article 4(6): a non-professional user who "simply launches" a compliant system according to its purpose is not covered, but one who disregards instructions of use when choosing the area of operation or setting performance conditions may be.
Review, representative actions and transposition (Articles 5–9)
Article 5 requires the Commission, five years after the end of the transposition period, to review the directive and report to the Parliament, the Council and the European Economic and Social Committee, where appropriate with a legislative proposal. The report is to examine the effects of Articles 3 and 4 and "in particular, it should evaluate the appropriateness of no-fault liability rules for claims against the operators of certain AI systems, as long as not already covered by other Union liability rules, and the need for insurance coverage", taking account of effects on AI uptake, especially for SMEs. A monitoring programme is to collect data, with member states reporting by 31 December of the second full year after transposition and annually thereafter. The memorandum explains the deferral on the ground that AI-equipped products and services that could put rights to life, health and property at risk, and so might warrant strict liability, "are not yet widely available on the market".
Article 6 adds the directive to Annex I of the Representative Actions Directive (EU) 2020/1828, enabling collective consumer actions. Article 7 sets a transposition deadline of two years after entry into force; Article 8 provides for entry into force on the twentieth day after publication in the Official Journal; Article 9 addresses the directive to the member states.
Key claims
| Claim | Basis | Confidence |
|---|---|---|
| The proposal harmonises only disclosure of evidence and the burden of proof, leaving fault, damage and causation standards to national law | Arts. 1(1), 1(3)(d); recitals 10, 13 | high (primary text) |
| It does not reverse the burden of proof; both presumptions are rebuttable | Explanatory memorandum part 2; Arts. 3(5), 4(7) | high (primary text) |
| For high-risk providers, fault supporting the causality presumption is limited to breach of listed AI Act requirements | Art. 4(2); recital 26 | high (primary text) |
| Strict liability and mandatory insurance were deferred to a review five years after transposition | Art. 5(2); memorandum part 5 | high (primary text) |
| Targeted harmonisation would add 5–7% to relevant cross-border trade and EUR 500 million to 1.1 billion in AI market value | Commission estimate based on Deloitte economic study | low (single forecast, not tested) |
Relationships
- supports: EU AI Liability Directive (proposed, withdrawn) — primary text of the proposal
- depends-on: EU AI Act (Regulation 2024/1689) — definitions and the Article 4(2)–(3) fault triggers are taken from the AI Act proposal
- instance-of: AI Liability — fault-based harmonisation through disclosure and rebuttable presumptions
- related: AI and Tort Liability — tort doctrine as applied to AI
- related: EU Cyber Resilience Act — named in the memorandum as complementary
- related: European Parliament — the 20 October 2020 resolution requesting a civil liability regime
- related: Regulating Under Uncertainty: Governance Options for Generative AI (G'sell, 2024), Inside My AI Law & Policy Class 11: When AI Fails, Who Pays? (Farahany, September 2025) — secondary accounts of the proposal