Why the EU Withdrew Its AI Liability Directive

The AI Liability Directive was the half of the EU’s AI package that dealt with compensating people harmed by AI systems. It never became law. The Commission listed it for withdrawal in its 2025 work programme, and the reason given was procedural rather than substantive.

What the proposal contained

The proposal, COM(2022) 496 of 28 September 2022, was published on the same day as the revision of the Product Liability Directive and was designed to sit beside it. Where the PLD is strict liability for defective products, the AILD addressed fault-based claims under national law — the ordinary negligence action a claimant brings in a member state court. Its text is on EUR-Lex as CELEX 52022PC0496.

It did two things and only two things. Article 3 would have let a national court order a provider or user of a high-risk AI system to disclose relevant evidence about that system, backed by a presumption of non-compliance if the order was ignored. Article 4 would have created a rebuttable presumption of a causal link between the defendant’s fault and the model output, where the claimant had shown fault, and where it was reasonably likely that the fault influenced the output.

Neither created a new standard of care. The proposal deliberately left the substance of negligence to national law and attacked only the evidential asymmetry — the fact that a claimant cannot see the training data, the evaluation results or the logs, and so cannot prove causation in a system whose behaviour is not documented in public.

Its scope was borrowed rather than defined. The disclosure power and the strongest form of the causation presumption were tied to high-risk AI systems as the AI Act defines them, which meant the directive’s reach would move whenever Annex III moved. That borrowing was one of the proposal’s known weaknesses: a claimant harmed by a system that is not high-risk — most consumer-facing generative systems — would have got a weaker presumption or none. The proposal also carried a review clause requiring the Commission, some years after transposition, to assess whether strict liability rules and mandatory insurance for certain AI systems were warranted. That was the door to a more ambitious regime, and it closed with the rest of the text.

The withdrawal and its stated reason

On 11 February 2025 the Commission published its 2025 work programme, COM(2025) 45. Annex IV of a work programme is the list of pending proposals the Commission intends to withdraw, and the AILD appeared there. The reason recorded against it was that there was no foreseeable agreement between the co-legislators, with an indication that the Commission would consider whether a different proposal should be tabled or another approach taken. The document is at EUR-Lex, CELEX 52025DC0045.

“No foreseeable agreement” is the standard formula and it is worth reading as what it is: a statement about the state of negotiations in the Council and Parliament, not a conclusion that the problem the proposal addressed had gone away. The European Parliament’s own research service had, in a complementary impact assessment published in September 2024, argued the opposite — that the instrument should be kept and broadened, potentially into a regulation covering more than high-risk systems. The Parliament’s legal affairs committee objected to the withdrawal. The Commission proceeded.

What survives, and where

The two mechanisms did not vanish entirely; one of them was already duplicated. Directive (EU) 2024/2853, the revised Product Liability Directive adopted on 23 October 2024, contains both a disclosure power and a set of presumptions:

• Disclosure. Article 9 of the PLD lets a court order a defendant to disclose relevant evidence at its disposal, subject to proportionality and trade-secret protection, with a presumption of defectiveness under Article 10 if the order is not complied with.
• Complexity relief. Article 10 of the PLD allows defectiveness and causation to be presumed where the claimant faces excessive difficulties, due to technical or scientific complexity, in proving them — provided the claimant shows the product contributed to the damage and that it is likely the product was defective. The defendant may rebut.

Those are close cousins of AILD Articles 3 and 4. The difference is the scope they apply within, and that difference is the substance of what was lost. See the revised Product Liability Directive and AI software for how far its own scope reaches.

What was lost with it

The PLD is a product-defect regime. It compensates death, personal injury, damage to consumer property and destruction or corruption of non-professional data, caused by a defective product. It does not touch pure economic loss, it does not touch discriminatory treatment, it does not cover harm to a business, and it does not reach a service that is not supplied as a product.

The AILD would have applied its disclosure and causation tools to fault-based claims generally, which is where those harms actually live. A person refused credit or a job by a model, a company whose operations were damaged by a defective AI service, a claimant alleging a duty of care was breached in how a system was deployed — all of them are now back on unmodified national tort law, proving fault and causation without any right to see the system. That gap is the whole subject of who is liable when a high-risk AI system causes harm.

Status and what would change it

As at August 2026 there is no replacement instrument in force and no adopted successor proposal. The Commission has signalled that AI-related liability may be revisited as part of its wider digital simplification and evaluation work rather than as a standalone directive, which is a signal and not a legislative act.

Three things would change the picture: a fresh Commission proposal, a member state legislating its own AI liability presumptions nationally — nothing prevents that, and it would fragment the position the AILD existed to harmonise — or a national court applying existing tort law in a way that shifts the evidential burden without a statute. Watch for the second. Divergent national rules are exactly the outcome the proposal’s impact assessment predicted from doing nothing.