Software becomes a product: what the new EU liability rules mean for you

A postmark for software

Software has never had a sell-by date quite like this. From 9 December 2026, if you place software on the EU market—AI systems included—it is a product, and you are liable for the harm it causes.^[1]

Think of it like this: a postmark proves when a letter was sent, and nobody expects the sender to trust the postal service to keep that record straight. The new liability rules give claimants tools to get at the evidence, and if you cannot show what your software did on a given day, the court may presume it was defective. That is the rule.

What counts as a product

The new Product Liability Directive treats standalone software, AI systems and digital services as products.^[1] Liability is strict: the claimant does not have to prove negligence, only that the product was defective, that they suffered harm and that the defect caused the harm.

It covers death and personal injury, damage to property not used exclusively for professional purposes, and destruction or corruption of data not used for professional purposes.^[1] Pure business losses are out.

Joint and several liability runs across the supply chain. If a product has multiple liable parties—say, a manufacturer and an importer—the claimant can go after any one of them for the full amount.^[1]

Evidence disclosure and burden-of-proof presumptions

Here is where it gets practical. Under certain conditions, a court can order the defendant to hand over relevant evidence.^[2] If the defendant does not comply, or if the claimant shows it is excessively difficult to prove a fact, the court may presume that fact is established.^[2]

Say someone claims your AI system made a faulty credit decision. If you cannot show which model version ran, what data it used and what it actually decided, the court may presume the system was defective and caused the harm. The presumption is rebuttable, but the burden has shifted.

That is new, and it makes your records everyone's business.

The Swedish transposition

Sweden's draft Product Liability Act, presented in August 2026, proposes that the new rules enter into force on 9 December 2026.^[3] The government is following the directive closely, and the timeline is tight. Other member states are doing the same, with varying speeds and slight differences in wording, but the substance is aligned.^[2]

The key date is when the product is placed on the market. Anything released after 9 December 2026 falls under the new rules, regardless of when development started.

What this means for you

If you build, sell or deploy software, ask yourself these questions now:

• Could you show which version of your software was running on a specific date, in a specific installation?
• Could you show what decisions it made, what data it used and which model version it relied on?
• Could you do that two years after the event, when someone files a claim?
• If a court ordered you to disclose that evidence, could you hand it over without having to trust that nobody in your organisation has quietly edited the records?

Version control is built for collaboration, not for evidence. It tells you what changed and when, but it does not prove that the record itself has not been altered. A Git log is useful until someone rewrites it.

That is the gap.

Sealing software versions and decisions

A digital witness seals a fingerprint of each release and each decision record as it happens. The content stays with you. The fingerprint goes onto a chain where any later change would show, even one made by someone with administrator rights.

Say you release version 2.4.7 of a loan approval model on 15 March 2027. You seal the build fingerprint, the release notes and the test results at release. On 3 April, the model declines an application. You seal a record of that decision: the applicant reference, the model version, the score and the reviewer.

Six months later, the applicant disputes the decision. You share the sealed records. They calculate the fingerprints themselves and check them against the sealed ones on WitniumChain. If they match, the records are unchanged. If they do not, someone has edited them, and that shows.

The witness does not prove the decision was fair, lawful or correct. It shows that the record is the one written at the time. That is enough to answer the first question a court will ask: what did the system actually do?

You can seal releases and deployments as part of your build pipeline, through the REST API, the SDK or the MCP server. See the documentation for how. The first 2,000 witnesses a month are free.

How it would work

1. At release, seal the build fingerprint, release notes and test results.
2. At deployment, seal a record of which version went where and when.
3. When a decision is made, seal the input reference, model version, output and reviewer.

Anyone you share the records with can check the fingerprints themselves.

What to watch

The new Product Liability Act is expected to enter into force in Sweden on 9 December 2026.^[3] The EU's AI Act applies high-risk requirements to Annex III systems from 2 December 2027, including logging duties under Article 12.^[4]

If your software makes decisions about people, or controls something that could harm them, both deadlines matter. The liability rules ask you to show what your product did. The AI Act asks you to log it. Putting the two together, the question is whether you can show that the log is the one written at the time.

That is the evidence question, and it is worth answering early.