Picture a cosmetic clinic in Germany. Nice website, tidy booking form, and a little chat window in the corner that answers questions at 11pm so the receptionist doesn't have to. Someone asks it whether the two directors are specialists in plastic and aesthetic surgery. The bot says yes. Asked again, it adds a couple more titles for good measure.
Lovely. Helpful. Fast. And completely made up.
That's not a hypothetical, by the way. That's the case. As tww.law laid out, the Higher Regional Court of Hamm ruled on 12 May 2026 that the clinic was on the hook for what its chatbot said. The directors were doctors, but they hadn't completed the training for the title the bot handed them. Two of the other titles don't exist as medical specialties at all. And the bot just invented them.
The website itself never made the claim. Only the chatbot did. But the court didn't care.
The defence that sounds reasonable until you say it out loud
The clinic's lawyers tried the argument most of us would try. The chatbot works on its own. Nobody wrote that answer. It's basically a third party. And everyone knows AI makes mistakes, so nobody takes it seriously anyway.
Here's how Triple A Digital summarised the judgment: the court went as far as assuming the bot really did run autonomously, and ruled against the clinic anyway. And the reason is almost annoyingly simple. The company set the bot up, decided what it was there to do, and could change it whenever it liked. That's enough. The chatbot isn't a third party. It's a tool, and you answer for your tools.
So here's the bit I enjoyed most. On the "everyone knows AI gets things wrong" point, the court said there's no such general experience. A lot of people trust a computer's answer more than a human's, because machines are supposed to be less error-prone than people. And if customers genuinely didn't trust the bot, the court noted, why would the clinic have put it on the website in the first place?
Think about that for a second. The trust is the reason you installed the thing. You can't then claim nobody trusts it when it drops a clanger.
The fix that was too late
This is the part that should make you wince. After the warning letter landed, the clinic switched the bot off and fixed it. A line in the prompt telling it to stay neutral whenever someone used the word "specialist", plus a keyword filter on the way out. Job done in what sounds like an afternoon.
It lost anyway.
Why? Well, the fix proved the point against them. If a prompt and a filter could stop the false titles after the complaint, a prompt and a filter could have stopped them before it. The court put it as a standard: the question was an obvious one, because anyone thinking about cosmetic work wants to know whether the practitioner is a specialist. And the fix was cheap. So the clinic should have tested for it before launch.
What was easy afterwards, you could have done beforehand. I'm stealing that line.
Not just a German problem
Now, before you say "well, I'm not in Germany, I'm not a cosmetic clinic": fair. It's a regional appeals court, not the top one, and the judgment is now final according to tww.law. It doesn't bind a UK or Australian court.
But it rhymes with the earlier one you might remember. Air Canada's bot told a grieving customer he could claim a bereavement fare after he'd travelled, with a link to a policy page that said the exact opposite. The airline suggested the bot answered for itself. A Canadian tribunal didn't buy it and awarded the customer CAD 812.02. It's small money. But a very big signal for the idea that "the bot said it" isn't a defence.
But two courts, two countries, same shrug at the same excuse. If you run a chatbot, a voice agent, or an AI that drafts replies to customers, assume the same logic lands on your desk one day.
What I'd do this week
Look, I don't think the answer is to rip out your chatbot. Mine are doing proper work. The answer is to treat the bot like a new hire on their first day, who is confident, fast, and has never read your policies. Here's the checklist I'd run.
1. List the questions where a wrong answer costs you money or trust. Qualifications, prices, warranties, delivery times, refunds, anything medical or legal. Write them down. That's your risk list, and it's usually under a dozen questions.
2. Ask the bot those questions yourself. Not once. Ask each one three or four different ways, including the slightly cheeky versions. The clinic's bot gave the false answer on a second ask, which is exactly when most of us would have stopped testing. Screenshot the answers.
3. Give it a short list of things it must never state as fact. Titles, guarantees, prices, anything regulated. Tell it to hand over to a human instead. Then add a filter on the output side as a second net, because prompts alone slip.
4. Link every factual answer to the real page. Air Canada's bot even did that, and the page contradicted the bot. If you can't point to the source for an answer, the bot shouldn't be giving it.
5. Read a sample of the transcripts every week. Ten minutes. You'll find the odd thing long before a consumer group does. This is the step everyone skips and the one that would have saved this clinic.
6. Don't lean on the disclaimer. "Answers may contain errors" sounds like cover. The writeup I read points out the court never said whether a notice like that would have changed anything, and the reasoning looks at the trust the bot creates, not the fine print under it. I wouldn't bet the business on a footer.
The bit that actually matters
Nobody in that clinic was being dodgy, as far as the reporting shows. Nobody sat down and decided to lie about qualifications. A tool they trusted said something confident and wrong, and because it spoke from their website, in their name, it counted as them.
So that's the whole lesson for the rest of us. When you hand a customer-facing job to an AI, you haven't handed away the responsibility. You've just given it a very fast mouth. So give it boundaries, test the awkward questions before launch, and read what it says.
If something's easy to fix after you get caught, it was easy to fix before.
Sources:
- Hamm Higher Regional Court: Companies Are Liable for False Statements Made by Their AI Chatbots, tww.law, June 2026, updated September 2026
- When your chatbot gets it wrong, you said it, Triple A Digital, September 2026
- Air Canada's Chatbot Invented a Procedure, Not a Price, UGO.IT, October 2026
Want to see how this could work in your business? Book a call and let's talk about where you're at and what's possible. Bring the chatbot, we'll interrogate it together.
Brewed by Steven, poured by Viktor
About Steven Tann: Steven helps business owners build systems that run themselves using AI. After 10+ years helping 7,000+ businesses and building his own autonomous operations, he's the bloke who actually does it, not just talks about it. Find out more at steventann.com.