Our customer-service chatbot sometimes gives answers that aren't true, even though we only fed it correct information. Courts in Germany and Sweden ruled on chatbot answers in 2026. Does EU consumer law treat a chatbot's wrong answer as the company's own statement, and what should we keep so we can show what the bot actually said?

What should I check, step by step?

1. Treat every customer-facing bot answer as a statement your business makes. Neither court accepted that the bot acted on its own or that a correct page elsewhere fixed a wrong answer in the chat.

2. List the questions customers will predictably ask about you: qualifications, licences, prices, contact routes, cancellation and refund rights. The Hamm court said a question about the doctors' specialist titles was foreseeable, so the clinic should have planned for it.
3. Test those questions against the live bot before launch and after every change to the model, prompt or knowledge base, and keep the test answers with the date and configuration version.
4. For claims you must never get wrong, add a fixed answer or a filter. After the warning letter, the clinic added a prompt instruction to answer neutrally about specialist titles and a keyword filter on the word, which the court took as proof it could control the bot.
5. Record each production answer (or at least its hash) with the time, the channel and the model, prompt and knowledge-base version, so you can show exactly what a customer was told.
6. When a complaint or warning letter arrives, preserve the conversation and the configuration that produced it before you change anything, then fix the answer.
7. Don't rely on fixing the answer later to end the matter. Hallon had changed its bot's answer, and the court still issued the ban; the clinic deactivated and reworked its bot, but because it didn't sign a cease-and-desist declaration, the court still found a risk of repetition.
8. Check your vendor contract separately. Under these rulings the customer-facing claim is yours whichever company built the model.

What did the German and Swedish courts decide about chatbot answers in 2026?

Hamm: a consumer association sued a clinic offering minimally invasive beauty treatments over answers from the AI chatbot on its website, which answered questions in real time and offered appointment booking after each answer. In April 2025 the bot said the two managing doctors were specialists in plastic and aesthetic surgery, and in other chats called them specialists in titles that don't exist. Neither had completed that specialist training. The court ordered the clinic to stop using those descriptions, with a penalty of up to EUR 250,000 per breach, and to pay the association's EUR 260 warning-letter costs. Because the questions were new, it allowed an appeal to the Federal Court of Justice; the state court database now lists the judgment as final.

• Sweden: the Consumer Ombudsman took Hallon to court over its hard-to-find customer service number and its chatbot's answer, "We are a digital operator, which means we don't offer customer service by phone." The court found the hidden number lawful, since customers had it in their contract terms, but held the chatbot answer misleading because existing customers could think phone service no longer existed, and so put off cancelling a subscription or disputing a bill. It banned that answer "or a similar answer" when a consumer asks for the number, under a SEK 500,000 conditional fine. The Ombudsman has appealed and is waiting to hear whether the Patent and Market Court of Appeal will take the case.
• Neither case needed a new AI law. Both applied ordinary rules on misleading commercial practices, which in the EU come from the Unfair Commercial Practices Directive.

Is giving a chatbot only correct data a defense if it still says something false?

The clinic argued that the bot worked autonomously on statistical probabilities, wasn't controlled answer by answer, and that nobody could tell whether the wrong answer came from the question, the training data or missing safeguards. The court accepted that it might not have been involved in each answer, but said the clinic set the bot's framework and could steer it, which the quick fix after the warning letter showed. That control was enough to make the answers the clinic's own.

• Because the answers were the clinic's own act, the court didn't apply the softer duty-of-care rules used when a business provides a platform that third parties misuse. The bot isn't a third party, and no other party would answer for its statements if the clinic didn't.
• The court also noted that a business deploys a customer chatbot because customers trust its answers, so it can't then argue that those answers carry no weight.

What records should you keep to show exactly what your chatbot told a customer?

Both cases turned on the exact words the bot used, and in Hamm on the date it said them. A business that keeps each answer with its time, channel and configuration can show precisely what was said, to whom and under which setup, and can show when a fix took effect.

• Keep the test runs too. A dated set of answers to the predictable questions, rerun after every change, shows you checked what the bot actually tells people rather than what it was meant to say.
• Store hashes of the conversation in a log you can't quietly edit, and keep the full text in your own system under your retention policy, so a later dispute can be checked against what was recorded at the time.