European Union - AI Liability Directive

Overview

The Proposal for a Directive on adapting non-contractual civil liability rules to artificial intelligence (AI Liability Directive, COM(2022)496) was presented by the European Commission on 28 September 2022 and aimed to address specific evidentiary and causation challenges arising from harmful outcomes involving AI systems. The draft sought to facilitate access to remedies by introducing a rebuttable presumption of causality and targeted disclosure powers for national courts, while attempting to coordinate with the EU's broader AI regulatory architecture, including the AI Liability Directive proposal and the Artificial Intelligence Act. Following protracted negotiations the Commission listed the proposal for withdrawal in its 2025 Work Programme (Annex IV) because it considered that no foreseeable agreement could be reached; the Commission later confirmed withdrawal after consultation with co-legislators. The proposal therefore remains a key reference point for future liability policy while not being advanced as binding law.

Definitions

The draft directive used several core definitions grounded in the AI Act and EU liability practice: "AI system" (broadly linked to the AI Act's definition covering systems that generate outputs such as content, predictions, recommendations or decisions), "operator" (the natural or legal person who has the power to exert control over an AI system), "provider" (the person who makes an AI system available on the market), and "damage" (material or non-material harm recognized under national tort law). The proposal also distinguished between harm caused by a physical product (where product liability rules may apply) and harm involving autonomous or software-based AI outputs, and it clearly addressed evidentiary concepts such as "information under control" and "relevant evidence" to be disclosed by defendants or third parties on court order.

Governance and Institutional Framework

The AILD would have operated primarily through national courts, with EU-level coordination. The Commission acted as proposer while Member States' competent authorities and national judiciaries would have implemented procedural and substantive elements. The proposal referenced existing EU instruments and institutions — notably the European Commission DG for Justice and Consumers, the European Parliament's Legal Affairs Committee (JURI) and the Internal Market and Consumer Protection Committee (IMCO) — as central actors in the legislative and oversight process. Enforcement and remedies were to be provided through civil procedures in national courts rather than through an EU administrative enforcement mechanism; the directive proposed harmonised rules on evidentiary presumptions and disclosure powers while leaving damages calculation and sanctions to national law, subject to principles of equivalence and effectiveness under EU law. The text contemplated safeguards for trade secrets and confidentiality when courts ordered technical disclosures.

Key Focus Areas

The principal substantive elements of the draft directive included (1) a rebuttable presumption of causality to alleviate the claimant's burden of proof where an AI system was implicated and evidence remained under the defendant's control, (2) explicit powers for national courts to order disclosure of relevant evidence including logs, training or design information and other technical documentation, (3) safeguards balancing access to evidence with protection of trade secrets and intellectual property, (4) definitions and scope rules clarifying when the directive complemented existing product liability and general tort regimes, and (5) coordination clauses to align the directive with the AI Act and with EU data protection rules, notably the General Data Protection Regulation (GDPR). Stakeholder debate focused on scope (should general-purpose models be included?), the exact triggering conditions for the causation presumption, the handling of trade secrets and security-sensitive information, whether to opt for harmonising regulation rather than directive, and the potential economic impact on innovation and SMEs.

Implementation Framework

Implementation would have required Member States to adopt national procedural and substantive measures to operationalise the directive's instruments. Key elements included: enabling domestic civil procedure rules to support court-ordered disclosure; creating mechanisms for secure and confidential handling of technical materials disclosed in litigation; training judges and court staff in technical basics of AI to handle complex evidentiary orders; ensuring compatibility with national damage calculation regimes; and setting proportionality safeguards (so that disclosure orders are limited to material evidence and do not unduly hamper business interests). The Commission envisaged guidance and exchanges of best practice among Member States, but left the exact implementation architecture (which national body performs what role) to Member States under the directive's provisions.

Monitoring and Evaluation

Under the draft, monitoring would have combined national court reporting and Commission-level stocktaking. The Commission expected to evaluate the directive's application over time and to collect data on litigation patterns, frequency and outcomes of disclosure orders, the use of the rebuttable presumption and any impacts on cross-border cases and market functioning. Interaction with the AI Act's conformity and market surveillance activities would have been monitored, and the Commission anticipated commissioning studies or impact assessments to assess whether the directive achieved greater access to remedies without creating excessive compliance costs or fragmentation.

Penalties, Liability, and Appeals

The directive focused on civil liability remedies: victims could pursue compensation through national courts. The proposal did not itself introduce administrative fines; instead it aimed to facilitate civil redress by easing proof hurdles and ordering evidence disclosure. Remedies included compensation for material and non-material damage, legal costs, and court-ordered disclosure. National courts would determine fault or strict liability as applicable under domestic law, subject to the directive's evidentiary rules. Appeals would follow national procedural law. The directive included measures to protect trade secrets in disclosure proceedings but did not create a separate EU-level enforcement agency or criminal sanctions.

Relationship to Other Instruments

The draft was explicitly designed to sit alongside, and to be coordinated with, existing EU instruments. It referenced the AI Act for definitions and the risk-based approach to AI; existing product liability rules (including the Product Liability Directive under revision); national tort law; the GDPR; and the Rome II Regulation on non-contractual obligations. Parliamentary and academic analyses recommended considering a more comprehensive software liability regulation to avoid fragmentation and to better align rules for general-purpose AI and software generally.

International Alignment

The proposal acknowledged the international dimension: AI systems and their providers are often global, so cross-border enforceability, mutual legal assistance for evidence and compatibility with third-country laws (e.g., trade secret regimes, data protection frameworks) matter. The draft sought to respect international obligations and to align with transatlantic and multilateral discussions on AI governance. The Commission and stakeholders noted that divergent national liability regimes risked creating trade frictions; conversely, overly prescriptive EU rules could affect competitiveness with third-country developers. The withdrawal left the EU to engage multilaterally while using the AI Act and other tools to promote interoperability and international norms.

Implementation Timeline

Sources and References