The AI Liability Act was withdrawn — here's what applies now
If you've been waiting for the AI Liability Act to pass to find out how liability for AI-caused harm gets allocated, here's the news: it was withdrawn almost a year ago, and nobody has proposed a replacement. Here's what actually fills the gap.
What the AI Liability Act was
The European Commission presented the proposal for an AI Liability Directive (AILD) on 28 September 2022, as a companion piece to the AI Act. Its goal was to adapt non-contractual civil liability rules to harm caused by AI systems, making it easier for victims to prove causation — particularly difficult with systems that function as black boxes. The European Parliament went as far as adopting it in March 2024.
Why it was withdrawn
On 11 February 2025 the Commission announced in its 2025 Work Programme that it was withdrawing the proposal, citing a lack of foreseeable agreement among the institutions on a final text. The official withdrawal was published in the Official Journal on 6 October 2025. MEP Axel Voss, the file's rapporteur, described the decision as caving to Big Tech pressure. The Commission said it would assess whether to table a revised proposal or a different approach — as of mid-2026, nothing has been announced.
This leaves a real gap: the AI Act defines what a high-risk system is and what documentation it requires, but it doesn't say what happens when that system causes harm and someone files a claim.
What fills the gap now — three pieces
1. National civil liability law
It never went away. Each Member State continues to apply its own non-contractual liability rules — generally fault-based. The problem the AILD was meant to solve — the difficulty of proving fault in opaque systems — remains unsolved at the EU level, so the outcome still varies by country.
2. The revised Product Liability Directive (2024/2853)
This one was actually adopted, and it's the piece that genuinely changed the playing field. Directive (EU) 2024/2853, published on 18 November 2024 and in force since 8 December 2024, modernizes the strict (no-fault) liability regime for defective products that had gone unchanged since 1985 — and for the first time explicitly includes software and AI systems as "products."
Three things make it especially relevant if you deploy AI:
- Component liability: if your AI system is integrated into another company's product (a model inside a medical device, a chatbot inside a third-party platform), you can be held jointly liable with the final manufacturer.
- Disclosure obligation for technical documentation: if a national court finds that a claimant has plausibly shown harm was caused by a defective product, it can order the manufacturer to disclose relevant technical documentation. This is exactly where the documentation the AI Act already requires comes into play.
- Claimant-friendly presumptions: in certain cases, the law presumes the product was defective or that it caused the harm, shifting part of the burden of proof.
Member States must transpose it into national law by 9 December 2026, and it will apply to products placed on the market after that date.
3. The AI Act as evidence
Although the AI Act doesn't regulate civil liability directly, its documentation becomes the evidence used in litigation. Risk classification, Annex IV technical documentation, usage and human oversight logs — all of that is exactly the kind of "relevant technical documentation" a court can compel under the new Product Liability Directive. National courts are already citing the AI Act's definitions and classifications when assessing AI liability cases, even though the AILD never came into existence.
What this means in practice
The practical takeaway is simple, if uncomfortable: the documentation the AI Act already requires isn't just for AI Act compliance — it's, in effect, your defense file if you're ever sued over harm caused by an AI system. The more complete and verifiable that documentation is (system inventory, risk classification, Annex IV technical documentation, human oversight logs), the better position you're in if a court orders its disclosure — because the alternative, not having it, plays directly into the presumptions the new Directive grants claimants.
Sources
European Commission 2025 Work Programme (11 February 2025); official withdrawal published in the EU Official Journal (6 October 2025); Directive (EU) 2024/2853 on liability for defective products, EU Official Journal (18 November 2024).
Your AI Act documentation, now with one more reason to keep it current
Annex IV technical documentation and risk classification aren't just for the AI Act auditor — they're your defense file in a product liability claim.