We use cookies and similar storage that the site needs to work. With your permission, we would also like to use marketing cookies to understand which of our pages and campaigns interest you and, if we know who you are, to let our team follow up. You can change your mind at any time from Cookie settings at the foot of every page. Cookie Policy
If you build or integrate software or AI, you are a "manufacturer" with narrower defences. You answer for third-party components you build in, and shipping updates keeps the risk alive.
The exposure is uncapped civil damages, with the burden of proof tilted against you. There are no fines, but courts can order disclosure of technical evidence and in set cases must presume a defect.
The rules apply to products placed on the market after 9 December 2026, whatever national lawmakers do. Most member states are late, texts are diverging, and no separate AI liability law is coming.
What should you do now?
Map products against the 9 December 2026 cut-off. Flag upcoming releases or retrainings that could count as "substantial modifications".
Treat documentation as evidence. Keep AI Act technical files, logs and test records complete and retrievable. Gaps turn a disclosure order into a presumed defect.
Run a defensible security-update programme. Review in-market products for missing patches, set support lifecycles, and align with the Cyber Resilience Act.
Rework contracts and insurance. Consumer terms cannot cap liability, so focus on indemnities and recourse with model and API suppliers, importers and distributors. Confirm your cover extends to software, data loss and psychological harm.
Audit marketing claims. Claims about AI accuracy or safety set the safety users are "entitled to expect".
Who carries liability for software and AI, and which defences remain?
"Product" expressly includes software and digital manufacturing files, such as 3D-printing templates. Supply model does not matter, so SaaS is in scope, and AI Act providers are treated as manufacturers.
Related services and components count. A "related service" is a digital service a product needs to perform a function, such as a cloud service controlling a smart appliance. Components can be intangible, and a manufacturer is liable for those it integrates under its control. In practice, an AI integrator answers for third-party models and APIs it builds in.
So does substantial modification. Anyone doing it outside the manufacturer's control becomes a manufacturer.
Defectiveness is an objective test aimed at software. A product is defective if it does not provide the safety a person is entitled to expect, or that the law requires. Courts weigh:
presentation, design and instructions (AI marketing claims feed in here);
reasonably foreseeable use;
the effect of any ability to keep learning after release;
the effects of inter-connected products;
where you keep control after release, the moment the product left your control, not just the sale date;
safety requirements, including safety-relevant cybersecurity requirements;
recalls or other safety interventions.
A product is not defective merely because a better version exists. AI unpredictability is no excuse. Security vulnerabilities can make a product defective; a hacker's contribution does not reduce liability.
Defences are narrowed for software. The development-risk defence remains: the defect was undiscoverable given scientific and technical knowledge at the time. For software, that time runs until the product left your control. The defence that the defect arose after release does not apply, where within your control, to related services, software including updates, missing safety or security updates, or substantial modifications.
There is no standalone duty to update, but failing to patch creates the exposure. Member states may restrict the development-risk defence: Finland keeps its exclusion; Germany and Sweden retain it.
Contracts cannot shield you. EULAs and terms cannot limit liability to the injured person, but businesses can allocate cost through indemnities and recourse. The one express tool is narrow: an integrator may waive recourse against a micro or small software supplier.
The liability window is long. Claims must be brought within 3 years of awareness of damage, defect and liable party. Rights expire 10 years after market placement. That clock restarts on substantial modification and extends to 25 years for latent personal injury.
What does the exposure look like if there are no fines?
Exposure is civil, not regulatory. The PLD has no regulatory fines and no enforcement authority, only civil damages decided by litigation. Final appellate judgments will appear in a Commission database.
No fault, no floor, no ceiling. Claimants prove defect, damage and causation, not negligence. The old regime had a 500 ECU property-damage threshold and an optional cap of at least 70 million ECU (Germany applied €85 million). Neither survives; liability cannot be limited by contract or national law.
New heads of damage. Medically recognised psychological harm and loss or corruption of non-professional data now count, alongside death, injury and private property damage. These are novel risks for chatbots, companion apps and data tools.
Disclosure. If a claimant makes the claim plausible, the court can order disclosure of relevant evidence. Disclosure must be necessary and proportionate, and trade secrets are protected. This is new for civil-law courts with no tradition of discovery.
Presumptions are mandatory, not discretionary. Defect shall be presumed where:
the defendant fails to disclose as ordered;
the product breaches mandatory safety requirements meant to prevent the harm, such as AI Act or Cyber Resilience Act rules; or
the damage came from an obvious malfunction in reasonably foreseeable use.
Causation is presumed where the damage is typically consistent with the defect. Even after disclosure, a court shall presume defect, causation or both if the claimant faces excessive difficulty, especially from technical complexity, and shows they are likely. This one is aimed at AI.
All presumptions are rebuttable, but in practice the burden moves to you. Fines sit in related regimes such as the AI Act and Cyber Resilience Act, and breaching those feeds directly into a presumption of defect.
Who can claim, and from whom. Only natural persons, or those acting for them or subrogated to their rights; businesses cannot claim in their own right. A cascade ensures an EU-based defendant always exists:
The manufacturer of the product, and of a defective component.
If the manufacturer is outside the EU, the importer and the authorised representative.
If neither exists, the fulfilment service provider.
Fallback only: distributors are liable only if they fail to name an EU operator or their own supplier within one month of a request. Online platforms count as distributors only where Digital Services Act conditions are met, broadly where a consumer would think the platform itself supplies the product.
Liability is joint and several: a claimant can recover in full from any liable party.
Claims can be aggregated. Practitioners say claims fall within the Representative Actions Directive (EU) 2020/1828, so damages can be aggregated across a user base. Material losses are uncapped (non-material where national law allows). Add defence and disclosure costs, and insurance premiums practitioners report are rising.
What does the exposure look like if there are no fines?
What does the exposure look like if there are no fines?
On claimant, with disclosure and mandatory presumptions
Outer time limit
Old regime (85/374/EEC)
10 years
New PLD (2024/2853)
10 years, restarting on substantial modification; 25 for latent injury
Regulatory fines
Old regime (85/374/EEC)
None
New PLD (2024/2853)
None
When do the new rules apply, and how far has transposition got?
The switch-over is fixed by the Directive. The rules cover products placed on the market after 9 December 2026; earlier products stay under the 1985 regime. Some commentators date this a day differently; rely on the wording "after 9 December 2026".
National transposition is patchy. Status per a September 2026 practitioner survey; later developments are mostly unconfirmed.
Adopted: Hungary (December 2025). Croatia and Lithuania are reported as adopted in mid-2026, but sources differ on their dates and status.
Germany: The Bundestag passed its reform act on 8 October 2026. The Bundesrat step and the entry-into-force date are not yet confirmed.
Advanced or moving: Slovakia, the Netherlands and Italy. Trackers disagree on Belgium and Slovenia.
Early stage or no known movement: the rest, including France, Spain, Austria and Ireland.
Late transposition will not remove your exposure. Expect it to be widespread, but courts must still read national law in line with the Directive.
Texts are already diverging. Finland and Sweden have dropped "excessive" from the complexity presumption in their drafts.
No separate AI liability law is coming. The proposed AI Liability Directive was formally withdrawn (notice published 6 October 2025), with no replacement. The PLD, alongside national tort law, is now the main harmonised route for AI harm claims.
What is the Product Liability Directive?
The PLD sets EU-wide strict liability for harm caused by defective products. It replaces the 1985 Directive (85/374/EEC). It was adopted on 23 October 2024 and entered into force on 8 December 2024.
Pure information, such as the content of e-books or bare source code, is not a product. Free and open-source software supplied outside a commercial activity is excluded, but once built into a commercial product, that product's manufacturer can be liable.
Which dates matter?
Which dates matter?
Which dates matter?
Date
Event
Date
23 Oct 2024
Event
Directive adopted
Date
8 Dec 2024
Event
Entry into force
Date
6 Oct 2025
Event
AI Liability Directive withdrawal published
Date
9 Dec 2026
Event
Transposition deadline; new rules cover products placed on the market after this date