Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

Who Is Liable When AI Fails: The AI Act Meets the Revised Product Liability Directive

The AI Act tells you when an AI system may be used; it does not tell you who pays when one fails. The revised Product Liability Directive (EU) 2024/2853 does — and it was rewritten with software and AI squarely in mind.

Renomira Team · 11 min read
A wood-panelled courtroom with benches and judicial portraits
Photo: Johan Pafenols / CC BY-SA 4.0 via Wikimedia Commons

Nobody plans for the failure that triggers the claim. A CV-screening tool filters out a qualified candidate; a recommender steers customers towards a listing that never ships; a predictive-maintenance model misses a bearing failure and stops a line for a week. Somebody wants compensation — and the question is never your AI policy, but which legal route the claim takes. This article maps them: the revised Product Liability Directive (EU) 2024/2853, national fault law, your contracts, GDPR Article 82 — and the AI Act itself, a regulatory regime rather than a compensation regime. The AI Liability Directive proposed in 2022 was withdrawn in February 2025, which reshapes the analysis.

Software is a product now: Directive (EU) 2024/2853

The 1985 Product Liability Directive was written for physical goods; whether code could be “defective” was left to decades of argument. The revised Directive (EU) 2024/2853 settles it: software, including AI systems, is a product.

  • No fault requirement. The claimant proves defect, damage and causal link — never your negligence.
  • Liability follows the supply chain. Producer first; for non-EU systems, the importer or authorised representative; the supplier answers only as a last resort, with recourse.
  • Producer status can shift to you. Rebranding a model as your own, or substantially modifying a system for a purpose its provider never foresaw — the same trigger as AI Act Article 25.
  • The calendar mirrors the AI Act. Transposition is due by 9 December 2026, application from 9 December 2027, days after the Annex III high-risk regime starts on 2 December 2027.

What makes an AI system “defective”

Article 7 keeps the classic test — the safety level the public at large is entitled to expect — and enumerates what a court weighs:

  • How the product was placed on the market: documentation, warnings, and whether marketing claims match reality.
  • Compliance with Union law requirements relevant to the product: AI Act conformity is read directly into defectiveness.
  • The ability to continue learning or evolve after placement — written for models that drift in production.
  • Reasonably foreseeable misuse, and effects on other products.

Two errors to avoid. The standard is objective, not omniscient: a drafting assistant producing clumsy prose is not defective; one sold as generating verified citations that invents them may be. For a self-learning system the deployment period stays in view, so monitoring records are your evidence.

Disclosure and presumptions: the evidence balance tips

Claimants historically lost AI cases for want of proof: the data, logs and code sat with the defendant. The Directive attacks that asymmetry:

  • Article 10 — disclosure of evidence. Courts can order production of source code, algorithmic documentation, evaluation records and logs; repeated non-compliance supports adverse inferences.
  • Article 11 — rebuttable presumptions. Where a quality regime was in place — in AI terms, a completed conformity route — the court may presume defectiveness, presume causation, or both.

Then the interdependence rule: a party that deviates from Union requirements — AI Act obligations included — or falsifies documentation cannot rely on that conformity to resist the presumptions. A documentation gap does not merely risk a fine; it hands the other side a presumption.

The route that vanished: the withdrawn AI Liability Directive

The 2022 proposal (COM(2022) 496) targeted fault-based claims — cases where somebody must be shown to have done wrong — and would have harmonised disclosure orders and presumptions triggered when the defendant controls the evidence. In February 2025 the Commission withdrew it, reasoning that the revised Product Liability Directive, the AI Act and national law covered the ground.

  • No harmonised fault route remains. National tort law governs, and diverges sharply on duties, evidentiary thresholds and limitation periods. Damage outside the strict heads — pure economic loss, most non-material damage — must rest on national law, contract or GDPR.
  • On AI liability directive status 2027: withdrawn, not on the current legislative programme. Planning around its return means planning around a document that no longer exists.

The liability map: five routes, four questions

Liability route Legal basis Who pays What the claimant must prove
Strict product liability Directive (EU) 2024/2853, as transposed nationally Producer; importer or authorised representative for non-EU systems Defect, damage, causation — no fault; Arts 10–11 shift the burden
Fault-based tort National civil law Whoever owed a duty of care — deployer, vendor or integrator Duty, breach, damage, causation on divergent national standards
Contract SaaS, licence or maintenance agreement The counterparty, within agreed caps and exclusions Breach — unavailable to applicants, employees and customers
Data protection damages GDPR Article 82 The controller, and processor where liable Infringement plus material or sufficiently serious non-material damage
AI Act enforcement Regulation (EU) 2024/1689, Arts 99–100 The fined company — to the state, not the victim A regulatory breach; the fine is not compensation

The victim’s claim runs down the first four rows; the AI Act fine runs down the fifth and never crosses over. You can pay a €15 million fine and still owe compensation.

What the AI Act does — and does not — do for liability

  1. Compliance is not immunity. The AI Act sets a legal floor; Article 7 asks what the public is entitled to expect. A system can be fully documented and properly overseen, yet still defective.
  2. Breach is evidence. A missing oversight process or an undisclosed AI interaction is what a claimant needs for a fault claim — and forfeits your position on the Article 11 presumptions.
  3. The victim does not get your fine. Penalties fund the state; compensation runs on separate tracks.

GDPR Article 82 fills part of the gap: anyone suffering material or non-material damage from an infringement may claim against the controller. Recruitment is the standard case — the rejected applicant needs no contract with anyone. Non-material damage must reach a degree of seriousness, but distress and lost employment are routinely argued.

Contract or tort? Your SaaS terms — caps, exclusions of indirect loss, indemnities — bind only the parties who signed. Everyone affected downstream proceeds in tort or under the Directive: supplier liability for AI-generated decisions cannot be negotiated away in a contract the affected person never saw.

Three mid-market scenarios

AI-assisted hiring. A manufacturer deploys a licensed CV screener; a rejected candidate claims under GDPR Art 82 and national tort law. The deployer answers for the decision it made — controller liability does not transfer to the vendor — while the vendor answers only for defect in its software. The deployer’s protection is its oversight record: human review, bias testing, logs.

Recommendation errors. A marketplace recommender pushes buyers towards a seller who never delivers. Injury from a physical product is classic producer liability; harm from the ranking itself — mis-selling, discriminatory visibility — falls to the platform as deployer, in tort and under consumer law.

Predictive maintenance failure. A model misses vibration signatures, a drive fails, the line stops for five days. Two claims stack: the machine as property injured by defective software — strict liability, subject to the Directive’s ceiling for professional-purpose property — and the lost production, consequential economic loss better left to the maintenance contract and business-interruption insurance.

What your vendor contracts must contain

  • A stated AI Act role and conformity position per system, with documentation referenced.
  • Change notification for model updates and purpose changes — a purpose change can move producer liability onto you.
  • Indemnities that survive the liability cap for data-protection and AI-decision claims, defence costs included.
  • Insurance with AI extensions — check for an “algorithmic error” exclusion, the clause that voids most mid-market claims.

How to check your own position

  1. List every AI system you deploy and AI feature you sell — liability attaches per system and role.
  2. Name the producer for each, including whether a white-labelled or modified system makes you one.
  3. Map affected people outside privity: applicants, customers, employees — each is a tort or Art 82 claimant.
  4. Check your defectiveness evidence — documentation, misuse analysis, monitoring records — and your presumptions position: any deviation from AI Act obligations forfeits reliance on conformity.
  5. Read your caps and exclusions against realistic losses, not contract value, and confirm your insurance covers algorithmic error and regulatory defence costs.
  6. Record the analysis with dates — an undated conclusion is close to no conclusion.

Frequently asked questions

Does compliance with the AI Act mean we cannot be sued?

No. The AI Act creates obligations and fines payable to the state — no immunity from civil claims. A conformant system can still fail the safety level the public is entitled to expect under Article 7 of Directive (EU) 2024/2853, and affected individuals can still claim under GDPR Article 82 or national tort law. Compliance helps rebut presumptions; it never closes the courthouse door.

Who is liable when an off-the-shelf tool we buy makes the mistake?

Usually the producer, under strict product liability: defect, damage and causation, no fault required. But you stay liable for the decision you made — as GDPR controller, as deployer for breach of your duties, and in tort towards affected third parties. Rebranding the tool as your own, or substantially modifying its purpose, can shift producer status to you.

What happened to the AI Liability Directive — will it return?

The Commission proposed it in September 2022 and announced withdrawal in February 2025, reasoning that the revised Product Liability Directive, the AI Act and national law covered the ground. It is not on the 2026–2027 legislative programme, so fault-based claims proceed under divergent national tort law.

Can we recover from our SaaS vendor when its model fails?

Only as far as your contract allows. Caps, exclusions of indirect loss and short notice deadlines routinely defeat otherwise strong claims — and they bind only you, not the people your decisions affect. Where the software is itself defective, the strict route under Directive (EU) 2024/2853 runs independently of the contract from 9 December 2027.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

liability for ai systems euproduct liability directive ai softwarewho is liable when ai makes a mistakeai act and liability rulesdefective ai product compensation euai liability directive status 2027supplier liability ai generated decisions

Keep reading