The North Sea at night under a low sky, one distant light on the horizon

Field notes Governance

From 9 December, software is a product. Find out if you made one.

November 2026  ·  7 min read

Most AI compliance work in Europe is organised around the AI Act. The law that will decide who pays when an AI system hurts someone is a different one. The Product Liability Directive, Directive (EU) 2024/2853, replaces the 1985 regime. Member States must bring it into national law by 9 December 2026, and it applies to products placed on the market or put into service after that date.

Its central change for anyone building with AI fits in one line of Article 4: a product "includes electricity, digital manufacturing files, raw materials and software". Liability under the Directive is strict. An injured person does not have to show that anyone was careless.

What the Directive says

Recital 13 removes any doubt about AI and about delivery models. Software is a product "irrespective of whether the software is stored on a device, accessed through a communication network or cloud technologies, or supplied through a software-as-a-service model", and "AI system providers within the meaning of Regulation (EU) 2024/1689", the AI Act, "should be treated as a manufacturer".

The claimant still has to prove three things: that the product was defective, the damage, and the causal link between them (Article 10(1)). Defective means the product "does not provide the safety that a person is entitled to expect or that is required under Union or national law" (Article 7(1)). Among the circumstances a court weighs is "the effect on the product of any ability to continue to learn or acquire new features after it is placed on the market".

Proof is where the Directive leans towards the claimant. A court must order the defendant to disclose relevant evidence once a claim is plausible (Article 9). Defectiveness is presumed if the defendant does not disclose, if the product breaches mandatory safety requirements, or if the damage came from "an obvious malfunction" in ordinary use (Article 10(2)). And where technical or scientific complexity makes proof excessively difficult, a court must presume defectiveness, causation or both once the claimant shows they are likely (Article 10(4)). Recital 48 names the cases it has in mind: machine learning, and a link that "would require the claimant to explain the inner workings of an AI system".

There is no ageing out through updates. The usual defence, that the defect did not exist when the product shipped, does not apply where the defect is due to software, a software update, "a lack of software updates or upgrades necessary to maintain safety" or a substantial modification within the manufacturer's control (Article 11(2)).

What it does not say

It is easy to read all this as "every wrong AI output is now a defect". It is not, and planning on that basis will put effort in the wrong place.

Only natural persons can claim. Article 1 limits the Directive to "damage suffered by natural persons". A business that loses money because a supplier's model gave a bad answer has no claim under it. Its remedy is the contract.

Only three kinds of damage count. Article 6 lists them: death or personal injury, including "medically recognised damage to psychological health"; damage to property, except the product itself and property "used exclusively for professional purposes"; and "destruction or corruption of data that are not used for professional purposes". Pure financial loss is not on the list.

Information is not a product. Recital 13 says product liability "should therefore not apply to the content of digital files, such as media files or e-books or the mere source code of software". The system is in scope. Its text output, on its own, is not the product.

Other routes stay open. Article 2(4) leaves contractual liability, fault-based national liability and data protection law untouched. The proposed AI Liability Directive, which would have eased fault-based claims about AI, was withdrawn by the Commission in 2025. That leaves this Directive as the EU-wide no-fault route, with fault claims left to each Member State's own law.

The Directive does not make AI output a defect. It makes the business that puts an AI system in front of people answerable for whether that system is safe, without anyone having to prove fault.

Who counts as the manufacturer

This is the question a deploying business has to answer first, because the Directive's answer is wider than most expect. Under Article 4(10), a manufacturer is anyone who develops a product, has one designed or made, puts "their name, trademark or other distinguishing features on that product", or develops one "for their own use". Article 8(2) adds anyone who substantially modifies a product outside the manufacturer's control and then makes it available or puts it into service.

Read against an ordinary AI programme, that gives four positions:

You bought it and use it as supplied. The supplier is the manufacturer. Your exposure under this Directive is low, though your contracts and duty of care still apply.

You built a system on a supplier's model and offer it to people. You are the manufacturer of that system. The model supplier may also be liable as the manufacturer of a component (Article 8(1)(b)), but that does not reduce your liability.

You put your brand on a supplier's assistant. Article 4(10)(b) treats you as presenting yourself as its manufacturer.

You retrained, rewired or extended a system so that it does something its maker did not assess. If that changes the nature of the hazard or raises the risk, Article 4(18) defines it as a substantial modification, and you take the manufacturer's place for the modified product.

Where more than one operator is liable for the same damage, they are liable jointly and severally (Article 12). Article 15 says liability may not, "in relation to the injured person", be "limited or excluded by a contractual provision". Your liability cap with your supplier still governs what you can recover from them afterwards (Article 14). It does not change what the injured person can recover from you.

What to do before 9 December

Classify every AI system. For each one, record whether you bought it, built it, branded it or changed it. That classification decides whether you are a manufacturer, and it should be written down before anyone needs it.

Find the people it can reach. List where each system's output can lead to physical harm, harm to mental health, damage to someone's property or the loss of their personal files. A customer-facing agent that can change account settings, book appointments or give health information is a different exposure from an internal drafting tool.

Keep the evidence you would have to disclose. Article 9 disclosure and the Article 10(2)(a) presumption mean the record matters. Keep the model and version in use, the configuration, the evaluations run before release and the update history, tied to dates. A business that cannot produce them starts from a presumption against it.

Treat updates as a safety duty. Under Article 11(2), a missing update that safety required is not a defence. Agree with each supplier how security and safety updates reach you, and record when you applied them.

Read the supplier contract for recourse, not for limits. Check what you can recover from the supplier if its component caused the damage, and whether the contract waives that right. Article 12(2) lets a micro or small software supplier agree such a waiver with you.

Watch national transposition. The Directive sets the rules, but each Member State's implementing law decides the procedure, and Article 18 lets states keep or introduce liability even where the state of scientific knowledge could not have revealed the defect. Check the law in each country where your systems reach people.

The time limits are long. A claimant has three years from knowing of the damage, the defect and the liable operator, and the right expires ten years after the product was placed on the market or substantially modified, or twenty-five for latent personal injury (Articles 16 and 17). Records kept today are the defence for a claim in the 2030s.

How we help

This note describes the Directive. It is not legal advice, and your exposure depends on national law and on facts only your counsel can weigh. Every quotation above is from the Official Journal text, and anyone can check it against the source.

We help businesses answer the manufacturer question for every AI system they run, and build the record the Directive assumes exists: version, configuration, evaluation and update history, held on your side. If you could not say today which of your AI systems you are the manufacturer of, talk to us.

All field notes

Tell us what you are accountable for.

The useful first conversation is usually about the thing that worries you, not about our capabilities. If it turns out we are the wrong firm for it, we will say so.

One reply from a person who has read what you sent.

Colophon

  • Plate 06, the sea at night: generated, Azure OpenAI gpt-image-1.5, 3 November 2026.

Every plate on this page is a generated stand-in for commissioned photography. The prompt that made each one is embedded in its file. How this site was made.