“The chatbot did it” no longer works
The Higher Regional Court of Hamm has ruled that companies are liable for what their AI chatbot says. What helps now and what doesn’t.

In shortKurz gesagt
On May 12, 2026, the Higher Regional Court of Hamm ruled that companies are liable for what their AI chatbot says. That applies even when the AI hallucinates and nobody ever entered the false statement. Legally, the chatbot isn’t an independent third party. It’s a tool of the company. The ruling is final. So what the bot says, the company says. That makes the question “What is our bot actually allowed to claim?” a matter for top management, in technology, design and organization.
What was the ruling about?
The case is quickly told, and it’s instructive. The consumer protection organization Verbraucherzentrale NRW sued a provider of cosmetic treatments. On the provider’s website, an AI chatbot answered questions from prospective clients and offered an appointment after every answer. Asked whether the two managing directors were board-certified specialists in plastic and aesthetic surgery, the bot said yes. They didn’t hold that qualification. In response to further questions, the bot also named the titles “specialist in aesthetic medicine” and “specialist in aesthetic treatments”. Neither exists.
The company defended itself with an argument that seems plausible at first glance. The bot had only accessed a vector database with the company’s own website texts and FAQs, and nothing in it was wrong. The false statements, it said, came from the AI’s autonomous, statistical behavior and therefore couldn’t be attributed to the company.
The court disagreed. It granted the injunction in full, based on the ban on misleading commercial practices under Section 5 of the German Act Against Unfair Competition (UWG) (case no. 4 UKl 3/25, full text from the NRW judiciary, in German).
What exactly did the court decide?
At its core, this is about who the bot’s statements are attributed to. The answer affects every company with a chatbot in customer contact.
The Higher Regional Court of Hamm made clear that an AI chatbot is not an independent third party. It’s a technical means the company uses to communicate with customers. Its statements therefore count as the operator’s own commercial conduct.
Four points from the reasoning are especially important.
- “AI is a black box” is no excuse. It makes no difference that nobody can see from the outside how the AI arrives at an answer. What matters, according to the court, is that the operator decides what the bot is responsible for and has enough influence over the system.
- The fix is what sealed the company’s fate. After receiving a cease-and-desist letter, the provider added a prompt instruction and a keyword filter, and the bot stopped making false claims. For the court, that was proof that the company can control the bot.
- The safeguards should have been in place before launch. It was foreseeable that prospective clients would ask about specialist titles. And the fix was obviously easy. So it would have been reasonable to build it in before the bot first went live. Clean source data alone wasn’t enough.
- “Everyone knows AI makes mistakes” doesn’t count. The company argued that users check AI answers anyway. The court countered that many people actually trust computer-generated answers more, because they consider machines less error-prone than humans.
The court drew a parallel to a 2021 ruling by Germany’s Federal Court of Justice on a software-based contract generator. There, too, the court attributed the software’s output to the provider. So the direction isn’t new. It’s now being applied to generative AI.
The court left open what happens when users deliberately provoke false answers with leading questions. In the Hamm case, the questions were neutral.
Why is chatbot liability a design and governance issue?
Many people overlook this point. The obvious reaction to a ruling like this is to ask the legal department and put a disclaimer on the chat window. Both fall short.
A disclaimer isn’t a solution. The court did not comment on disclaimers directly. But it rejected the argument that users know AI makes mistakes and therefore check its answers. A notice like “AI can make mistakes” relies on exactly that argument. That’s why the Wettbewerbszentrale, Germany’s self-regulatory body for fair competition, assumes that generic notices like this don’t reliably protect against liability. And in practical terms: if the bot tells a customer that her contract covers a certain kind of damage and she relies on it, the small print in the chat window won’t help her much.
That puts the problem where it belongs: in the architecture and design of the system. A bot that stays on the right side of the law comes from technical and design decisions. Legal wording alone won’t get you there. The key questions are:
- On which topics may the bot phrase answers freely, and where does it answer only with fixed, approved texts?
- Where does it get its facts? A system that is tied to verified sources hallucinates less often. We explain how that works in our article RAG explained simply.
- What happens when the bot is unsure? Does it guess, or does it hand over to a human?
But the Hamm case also shows the limit of verified sources. According to the company, the bot only accessed its own, correct content, and it still hallucinated. Verified sources are necessary, but not enough on their own. Only a technical limit like the later prompt instruction with a keyword filter actually stopped the false statements.
That’s why hallucinations aren’t an abstract technical problem either. They’re a concrete business risk. We’ve written in detail about why language models will always hallucinate and how systems can stay reliable anyway.
In the end, the liability question is a question for conversation design. Whoever decides what the bot may and may not say also decides the liability risk. A recent study shows how much missing limits hold AI projects back: Why 74% of companies roll back their AI agents.
Checklist: What companies should do now
The ruling and expert commentary add up to a practical list. It doesn’t replace legal advice, but it shows where design and governance come in.
- Set topic boundaries. Decide which questions the bot may answer and which it may not. Sensitive topics such as qualifications, commitments, or legally and medically relevant information call for fixed, approved flows instead of freely worded AI answers.
- Allow only verified sources. The bot should derive its answers from verified information, not freely from the language model. That lowers the risk of made-up statements.
- Design escalation. When a request gets too complex, unclear or sensitive, the handover to a human has to work smoothly. It’s then a deliberate part of the conversation, not a dead end.
- Set up monitoring. Catch false statements early and stop them fast. The court assumes you can control your bot anyway. So you should do it before any damage is done.
- Don’t rely on disclaimers. A notice like “AI can make mistakes” replaces none of the points above. The court explicitly rejected the argument that users check AI answers anyway.
- Document everything. Record which safeguards you put in place before launch and how you check them on an ongoing basis. That way you can trace errors and show that you take responsibility seriously.
What does this mean for regulated industries?
For telecommunications, banking, insurance and healthcare, the ruling is no side issue. These are exactly the sectors where chatbots give information people rely on, for example about pricing plans, terms, coverage and claims. By the logic of the Hamm court, every one of those answers is a statement by the company.
The court stresses that particularly strict standards apply in healthcare. How strictly other courts will judge other industries remains open. But the logic of attribution doesn’t depend on the industry. If you give binding information via a bot in finance, insurance or telecommunications, you should assume the same standard applies.
That isn’t bad news. The ruling creates clarity and pushes the market in a healthy direction, away from bots clicked together in a hurry and toward systems with clear limits, verified sources and clean escalation. In a regulated industry, a bot that knows what it isn’t allowed to say is worth more than one with an elegant answer to every question.
Then there’s the EU AI Act. Its transparency obligations under Article 50 have applied since August 2, 2026. Since then, anyone operating a chatbot has to tell users clearly that they’re talking to an AI. The direction is clear. Responsibility can’t be handed off to the machine.
Conclusion: Responsibility can’t be handed off to the machine
“The chatbot did it” no longer works as a defense, and now that the ruling is final, that’s settled. If you put an AI in front of your customers, you’re responsible for what it says, just as for any other statement your company makes. That sounds like a risk, but above all it’s a design task. A chatbot with clear limits, verified sources and an honest path to a human is legally safer. And it’s the better chatbot.
We build systems like that. Together with you, we decide what the bot is allowed to say, connect it to your verified sources and design the handover to your team along the way. Using a bot in an area where wrong answers get expensive? Talk to us!
FAQ: Chatbot Liability
Are companies liable for the statements made by their AI chatbots?
Yes, according to the final judgment of the Higher Regional Court of Hamm dated May 12, 2026 (Case No. 4 UKl 3/25). The court considers the statements made by an AI chatbot to be the company’s own business conduct, even in cases of hallucinations.
Does a disclaimer in the chat window protect against liability?
You shouldn’t rely on that. The court rejected the argument that users are aware of AI errors and verify responses. Many people actually trust machines more than humans. A generic notice such as “AI can make mistakes” does little to change this. Concrete technical and organizational measures are necessary.
Can a company argue that the AI hallucinated on its own?
No. In the court’s view, the chatbot is a tool of the company and not an independent third party. The court also did not accept the argument that the bot only accessed its own correct content. The false statement was foreseeable and easy to prevent, so the company should have taken precautions before the launch.
Does this apply only to the healthcare industry?
No. The case involved a cosmetic clinic, and particularly strict standards apply in the healthcare sector. However, the logic behind the ruling applies to all industries. It is especially relevant wherever bots provide binding information, such as in finance, insurance, telecommunications, and law.
What specific steps should companies take now?
Set topic boundaries for the bot, allow only verified sources, ensure a smooth handoff to humans, continuously monitor the bot’s behavior, and document all protective measures. The focus is on design and governance. Legal language alone is not enough.
Is the ruling final?
Yes. Although an appeal to the Federal Court of Justice was granted, the ruling has become final. There is therefore no Federal Court of Justice ruling on this issue. Other courts are not bound by the decision, but it sends a clear signal.
QuoteZitat