The AI liability directive does not exist, and liability is not left without rules because of that
If you look for an "AI Liability Directive" in force, you will not find one: the Commission's proposal (COM(2022) 496) has been withdrawn. The Commission work programme for 2025, published on 11 February 2025, lists it among the proposals to be withdrawn with the reason "No foreseeable agreement" and adds that the Commission "will assess whether another proposal should be tabled or another type of approach should be chosen". The procedure page on EUR-Lex records the withdrawal on 6 October 2025. Whoever writes software with machine-learning components therefore has to ask a different question: if my system causes harm, which rules actually apply?
The answer has three layers: national liability law (contract and tort), Directive (EU) 2024/2853 on liability for defective products, which from 9 December 2026 treats software as a product, and the GDPR when the damage comes from an unlawful processing of personal data. The AI Act, the subject of the first article of the series, is not on the list for a precise reason that I explain below. This article says what the rules provide, not how to win a lawsuit.
What the withdrawn proposal would have done
It is worth knowing, because the ideas of the proposal still circulate in popular texts as if they were law. The proposal did not create a new right to compensation: it intervened on evidence in fault-based non-contractual liability claims. It provided two instruments, which I re-read in the text of the proposal.
- In Article 3, the power of national courts to order the disclosure of evidence about a specific high-risk AI system suspected of having caused damage, and a rebuttable presumption of non-compliance with a duty of care for whoever fails to comply with the order.
- In Article 4, a rebuttable presumption of a causal link between the fault of the defendant and the output of the system, under precise conditions: for the provider, proof of non-compliance with requirements such as data quality (Article 10), transparency (13), human oversight (14), accuracy and cybersecurity (15).
The link with the AI Act was direct: the technical requirements of the regulation became the yardstick by which the injured party could prove fault. That link does not exist today as a uniform rule. If you read in an article that "non-compliance with the AI Act gives rise to a presumption of liability", it is a description of the proposal, not of the law in force.
The three layers that apply today
The first layer is national. The defective products directive expressly leaves untouched the rights that an injured person has under national rules on contractual liability or on non-contractual liability on grounds other than the defectiveness of a product (Article 2(4), point (b), and Article 6(3)). In practice, a customer who has a contract with you can invoke non-performance; a third party can invoke tort, bearing the burden of proving fault, damage and causal link under the rules of their legal system. I do not describe Italian law in detail here: I have not verified it for this article and the national part requires a lawyer.
The second layer is Directive (EU) 2024/2853. It repeals Directive 85/374/EEC, which keeps applying to products placed on the market or put into service before 9 December 2026 (Article 21), while the new one applies to products placed on the market after that date (Article 2(1)). Member States must transpose it by 9 December 2026 (Article 22); I have not verified the Italian transposition. Recital 13 says that, for the purposes of no-fault liability, software is a product irrespective of the mode of its supply or usage: stored on a device, accessed through a communication network or cloud technologies, or supplied through a software-as-a-service model. It expressly names AI system providers within the meaning of the AI Act among the developers or producers of software to be treated as manufacturers. Free and open-source software developed or supplied outside the course of a commercial activity is excluded (Article 2(2)), as is information, such as the content of digital files.
The third layer is the GDPR: Article 82 gives any person who has suffered material or non-material damage as a result of an infringement of the Regulation the right to receive compensation from the controller or processor. If your system decides about people using personal data that was processed unlawfully, this route exists regardless of the AI Act.
And the AI Act? The consolidated text does not contain an article that grants a right to compensation: it provides for administrative fines (Article 99), a right to lodge a complaint with a market surveillance authority (Article 85) and a right to explanation of individual decision-making (Article 86). The word "compensation" does not appear in the text at all, and "damages" appears only once, in a criterion for amending Annex III (Article 7(2), point (k)), which excludes "claims for damages" from the redress measures considered. Infringing the AI Act therefore does not automatically mean having to pay compensation; it can, however, weigh as evidence of a defect or of fault, as you see in the next section.
Where the new directive helps the injured person and where it does not
The directive covers only certain categories of damage (Article 6(1)): death or personal injury, including medically recognised damage to psychological health; damage to, or destruction of, property, except the defective product itself and property used exclusively for professional purposes; destruction or corruption of data that are not used for professional purposes. A credit-scoring system that wrongly refuses a mortgage produces economic and perhaps discriminatory harm, but none of those categories: for that case there remain national law, Article 82 of the GDPR if there is unlawful processing, and administrative remedies. A driver-assistance system or a triage system that causes an injury is instead inside the perimeter.
Inside the perimeter, the directive shifts the burden of proof in a way that matters to developers. A product is considered defective where it does not provide the safety that a person is entitled to expect or that is required under Union or national law (Article 7), and among the circumstances to consider are the effect of any ability to continue to learn after the product is placed on the market, safety-relevant cybersecurity requirements and the moment when the product left the control of the manufacturer. On the evidence side:
- the defendant can be required to disclose relevant evidence if the claimant presents facts and evidence sufficient to support the plausibility of the claim (Article 9), with protection of trade secrets;
- defectiveness is presumed if the defendant fails to disclose, if the product does not comply with mandatory product safety requirements intended to protect against the risk of the damage suffered, or if there is an obvious malfunction during reasonably foreseeable use (Article 10(2));
- a national court presumes defectiveness or the causal link where the claimant faces excessive difficulties, in particular due to technical or scientific complexity, and demonstrates that it is likely that the product is defective (Article 10(4)), and the defendant may rebut the presumptions (paragraph 5).
The second item raises a question that the directive does not resolve: do the requirements of the AI Act count as "mandatory product safety requirements"? It is an interpretative question on which I know of no case law, and I do not present it as settled. What you can say with certainty is the practical consequence: if you cannot produce the documentation and the logs that a court asks of you, the first presumption is triggered without anyone having to prove anything about your model.
Two rules concern the life cycle of software. The exemption for a defect that did not exist when the product was placed on the market (Article 11(1), point (c)) does not apply where the defect is due, while the product is within the manufacturer's control, to software including updates, to a related service, to a substantial modification or to the lack of the updates necessary to maintain safety (paragraph 2). Translated: as long as you control the updates of your model, a defect introduced or left uncorrected by an update remains yours. The exemption for the state of scientific and technical knowledge that did not allow the defect to be discovered remains (point (e)). The time limits: a limitation period of three years from when the injured person became aware, or should reasonably have become aware, of the facts (Article 16), and an expiry period of 10 years from placing on the market, extended to 25 years for personal injury with latency (Article 17).
What changes for whoever builds
The thread that links the AI Act and the defective products directive is evidence. The AI Act requires providers of high-risk systems to ensure automatic recording of events (Article 12), to keep the logs for at least six months unless otherwise provided (Article 19) and the technical documentation for ten years (Article 18). The directive provides for an expiry period of ten years and presumptions that are triggered when you do not produce the evidence. Neither rule tells you how long to keep a log in order to defend yourself: that is a risk choice, to be justified case by case, and the coincidence of the ten years is not a rule.
To know which directive you are looking at, the date matters. The following listing is a function with three branches: it receives the date of placing on the market and returns the regime, with the exception of non-commercial free software.
"""Which defective-products liability directive applies?
Directive (EU) 2024/2853, Art. 2(1) and Art. 21; Directive 85/374/EEC for the rest."""
from datetime import date
DATA_NUOVA_DIRETTIVA = date(2026, 12, 9)
def regime(immesso_sul_mercato, software_libero_non_commerciale=False):
if software_libero_non_commerciale:
return "out of scope: free and open-source software supplied outside a commercial activity (Art. 2(2))"
if immesso_sul_mercato > DATA_NUOVA_DIRETTIVA:
return "Directive (EU) 2024/2853 (software is a product: Art. 4(1))"
return "Directive 85/374/EEC (keeps applying to products placed on the market before)"
for quando in (date(2025, 3, 1), date(2026, 12, 9), date(2026, 12, 10)):
print(quando.isoformat(), "->", regime(quando))
print("open-source library, non-commercial release ->", regime(date(2027, 1, 15), True))
# Output (python3 -I, Python 3.14.4):
# 2025-03-01 -> Directive 85/374/EEC (keeps applying to products placed on the market before)
# 2026-12-09 -> Directive 85/374/EEC (keeps applying to products placed on the market before)
# 2026-12-10 -> Directive (EU) 2024/2853 (software is a product: Art. 4(1))
# open-source library, non-commercial release -> out of scope: free and open-source software supplied outside a commercial activity (Art. 2(2))
The comparison immesso_sul_mercato > DATA_NUOVA_DIRETTIVA uses "greater than" and not "greater than or equal" because Article 2 says "after 9 December 2026" and Article 21 keeps the old directive for products placed on the market "before that date": 9 December itself is an edge case that the text does not resolve, and the output treats it under the old directive out of caution, not because the rule says so. The function answers only the question of which directive to look at: it does not establish whether there is a defect, nor whether the damage falls within the categories of Article 6. For every new version of the model, the question "is it a new product or a modification?" matters: a substantial modification makes the period run again (Article 17(1), point (b)), and whoever substantially modifies a product can become a liable economic operator (Article 8).
Four engineering habits make proof less costly, whatever the regime: model versions identifiable in every log, logs that include the relevant inputs and the outputs without superfluous data (the fifth article explains the conflict with minimisation), a history of updates with their rationale, and a procedure for releasing known security fixes. They are common-sense practices, not obligations under that name.
Frequently asked questions
If I use a model from a cloud provider, who is liable for the damage?
It depends on who the liable economic operator is. The defective products directive points to the manufacturer of the defective product and the manufacturer of a defective component integrated within the manufacturer's control, whose liabilities add up without prejudice to joint and several liability (Articles 8 and 12). As an integrator you are generally the manufacturer of the final product; your contract with the provider then governs recourse, within the limits of the law.
If the damage is caused by wrong input data, am I liable?
It cannot be said in the abstract. The directive considers defective a product that does not provide the expected safety, taking into account reasonably foreseeable use and the presentation of the product, including instructions; under national rules, fault matters. A system that validates its inputs and documents the checks is in a more defensible position than one that accepts any data.
My system is for internal use only: am I covered by these rules?
The directive applies to products placed on the market or put into service, and "putting into service" is the first use in the Union in the course of a commercial activity, whether in return for payment or free of charge (Article 4). An internal use within a company may fall under it, but I have not verified how Italian law qualifies it: it must be assessed case by case.
Is insurance for AI damage mandatory?
Neither the AI regulation nor the defective products directive provides for it for AI system providers; the only liability insurance the regulation imposes concerns notified bodies (Article 31). Insurance may nonetheless be a contractual or commercial choice.
Does the directive also cover pure economic loss?
No: Article 6 lists death and personal injury, property other than the defective product and not used only for professional purposes, and non-professional data. A purely financial loss is assessed under national rules or under Article 82 of the GDPR when its conditions are met.
What you can do now
You know that the AI liability directive is a withdrawn proposal and where to verify it; you can tell apart the three layers that really apply; you know which damage the new defective products directive covers, from when and with which presumptions of proof; and you know which design habits make a system defensible. In the next article you move from the system to the model: the obligations of providers of general-purpose AI models. For the ML governance side you can also read ML Governance: Compliance, Explainability and Ethics.
Sources
- Directive (EU) 2024/2853 on liability for defective products - recitals 13 and 14, Articles 2, 4, 6-12, 16, 17, 21, 22, consulted on 2026-11-13
- Proposal COM(2022) 496 for a directive on non-contractual civil liability for AI - Articles 3 and 4 (disclosure of evidence, presumption of causal link), consulted on 2026-11-13
- Procedure file 2022/0303(COD) on EUR-Lex - status "Proposal withdrawn", withdrawal of 6.10.2025 (CELEX 52025XC05423), consulted on 2026-11-13
- Commission work programme 2025, COM(2025) 45, Annex IV - the AI liability directive proposal among the planned withdrawals, with the reason given, consulted on 2026-11-13
- Regulation (EU) 2016/679 (GDPR), text published in OJ L 119 of 4.5.2016 - Article 82, consulted on 2026-11-13
- Regulation (EU) 2024/1689, consolidated text of 27.7.2026 - Articles 12, 18, 19, 31, 85, 86 and 99, consulted on 2026-11-13
The listing was run with Python 3.14.4, with identical output in two runs. Italian liability law and the national transposition of Directive 2024/2853 were not verified; the GDPR text is the one published in 2016, not a later consolidated version.

Commenti
Caricamento commenti...
Sign in to comment