Your government contracts just got more transparent

On 12 May 2026, the Higher Regional Court of Hamm held a cosmetic treatment provider liable for statements its AI chatbot made about the doctors who ran the business. The chatbot invented qualifications they did not hold, and the court held the company responsible.

The company’s position was the position most businesses instinctively reach: the chatbot said it, not us. The system’s output was not drafted by a human, not signed off, and emerged from probabilistic behaviour nobody specified.

The court rejected this. The test is not technical. It is organisational. The chatbot’s statements were attributable to the company because the system forms part of the company’s own business organisation. The chatbot is not a third party whose misconduct relieves the operator of responsibility.

This matters commercially because it disposes of the second defence as well: “We did not build the model, we licensed it.” That is a statement about your supply chain, not about whose business the system was operating inside when it spoke to your customer. A government contractor using Government Assist cannot escape liability by saying the system belongs to DSIT.

Most businesses running a customer-facing chatbot have a line of text somewhere near it: Answers without guarantee. May produce inaccurate information. Please verify independently.

The court addressed this directly. A general disclaimer of that kind does not defeat liability. What is required instead is actual technical and organisational control over the system.

This is not unique to German law. The Committee of Advertising Practice has stated that the UK Advertising Codes apply regardless of how advertising content is generated, and that disclosure is very unlikely to mitigate the harm caused by a fundamentally misleading message.

Two regulators, two jurisdictions, no coordination, same answer: a disclaimer sitting next to a false statement does not convert it into a true one. The disclaimer may manage customer expectations. It will not manage your exposure.

The temptation is to file this as a German case and move on. That would be a mistake. The equivalent UK analysis is less forgiving, not more.

Unfair commercial practices offences under the Digital Markets, Competition and Consumers Act 2024, in force since 6 April 2025, are strict liability, apart from the professional diligence limb. The trader’s state of mind at the time of the offence is irrelevant.

Separately, the CMA can now impose penalties of up to 10 per cent of global turnover without going to court.

So a UK business whose chatbot invents a qualification, a price, an accreditation, a delivery window, an eligibility criterion or a warranty term is exposed on the same attribution logic Hamm applied, without needing a German court to get there, and against a regulator that no longer needs a court either.

The court pointed toward the answer when it said technical and organisational control is what is required. In practice that means five things:

• Log and retain the outputs, so that when a complaint arrives you can establish what was actually said rather than reconstructing it.
• Test before release against the questions that carry regulatory weight, qualifications, accreditations, warranties, delivery windows, eligibility criteria, rather than the ones about opening hours.
• Apply version control to prompts and models, with a record of what changed and when, because a model update you did not initiate can change behaviour you had already tested.
• Build a working escalation route to a human for the categories where a wrong answer is expensive.
• Name an owner, because a control nobody is accountable for is a policy, not a control.

Here is the position most businesses are in without having noticed.

Almost every conversational AI vendor agreement currently in the market excludes liability for model output. It is close to standard. The result is that the operator carries the whole of a strict liability consumer law exposure, created by a system it did not build, cannot fully inspect, and cannot prevent from changing.

That is a risk allocation nobody would accept if it were written down in those terms. It usually is written down, in the exclusions, and it usually is accepted, because the exclusions get less attention than the price.

For government contractors, this matters acutely. You are delivering to a client whose own regulatory obligations now flow through to you, and your vendor agreement passes the risk straight back to you.

In your government client agreements:

• Allocate responsibility for AI disclosure explicitly, specifying who confirms AI use and who ensures disclosure before publication.
• Specify review timelines before publication, not after. The client carries liability for content it approves; you carry liability for content you deliver.
• Include indemnities covering reputational harm from undisclosed or inadequately reviewed AI-generated content, because a false statement attributed to the government becomes your liability if you knew or should have known.
• Establish data-protection protocols for sensitive materials processed through third-party AI systems. The DSIT framework does not displace data protection obligations.

In your vendor agreements:

• Output logging should be a contractual deliverable with a defined retention period and your right of access, not a feature the vendor may deprecate.
• You should have advance notice of material model or prompt changes, because otherwise your pre-release testing has a shelf life you cannot see.
• There should be a testing obligation before any material change is pushed to production.
• And the indemnity needs to reach regulatory and consumer law claims arising from model output.