A clinic's chatbot invented medical titles, and a court said they were the clinic's own words

What happened

A clinic for aesthetic medicine in Recklinghausen, Germany, put an AI chatbot on its website. Visitors could ask questions and book appointments. When asked, the bot said the two doctors behind the company were specialists in plastic and aesthetic surgery, in aesthetic medicine and in aesthetic treatments. Those were not titles the doctors held, and two of the names do not exist as specialist titles at all.

The consumer organization Verbraucherzentrale NRW sent a warning letter and asked for a signed promise to stop. The company switched the bot off but did not sign, so the organization sued. On 12 May 2026 the Higher Regional Court Hamm ruled for the organization (case 4 UKl 3/25) and ordered the company to stop using those titles.

What the court said

As reported by the court's press release and by law-firm summaries:

The bot's answers counted as the company's own commercial statements. The court did not accept the argument that an AI system works statistically and is not steered by a person.
The chatbot is not a "third party" whose words are attributed to the company only in special cases.
The company was responsible even if it had set the bot up with correct data only.
The company could and should have foreseen that people ask about qualifications and that a bot might invent an answer.
Commentators, for example the Wettbewerbszentrale, read the ruling as saying that a generic notice such as "AI can make mistakes" does not reliably protect a company.

Keep the scope in mind. This was a competition-law case about misleading statements, it asked for an injunction rather than damages, and it applies German law. It is not final, so check where it stands before relying on it.

Why it happened: a design reading

A question about the company looked like an easy FAQ. "Who are you and what are you qualified to do?" feels harmless, but a wrong answer is a false claim.
The bot generated an answer instead of reading a fact. Nothing forced it to answer only from a maintained source, or to say "I cannot confirm that".
Correct data did not make the output correct. A model can still produce a fluent, plausible, wrong title.
Nobody was watching. The false answers were found by an outsider, not by the company's own review.

The same shape in support

Any assistant that answers about prices, warranty, delivery times, availability, licences, certifications or "what we guarantee" can make the same kind of claim. In many places a false statement in sales or marketing can be treated as the company's own, whoever or whatever wrote it. See module 18 for the wider picture and the Air Canada case for an earlier example.

Design lessons

Treat claims about your organization as a protected topic with its own source and its own owner.
Answer from the source or decline. "I cannot confirm that. Our team page lists the doctors' qualifications" beats a fluent guess.
Test the risky questions on purpose: credentials, prices, guarantees, availability. Run them again after every change.
Review real transcripts for claims about the company, on a schedule.
Do not count on a generic disclaimer to carry the risk.
Keep a record of what the bot said and be able to switch a topic off quickly.

How to build it better

Write a fact sheet about the organization: names, titles, services, prices, hours, guarantees. Give each fact an owner and a date.
Give the assistant a tool that reads the fact sheet, and make it refuse when the answer is not there (module 6, module 8).
Add an output check for risky words such as professional titles, "guaranteed", "free" and "certified", that blocks the answer or hands off (module 10).
Put a test set of company-claim questions into your evaluation suite (module 14, module 24).
Review a sample of transcripts every week and log what you fixed.

Try it

List ten things your assistant could say about your organization that would be false or misleading if wrong. For each one, name the source it should read and say what it replies when that source has no answer.