Chatbot liability for wrong information usually becomes a topic only after the fact: a customer sends a screenshot promising a refund or discount that your terms do not offer. Two decisions, one from Canada and one from Germany, back that intuition: neither treated the chatbot as a separate party. This article is not legal advice and does not predict how a court in your country would decide.
For a support team, the practical questions matter more: which topics carry the most risk, where the assistant takes its facts from, what it does when information is missing and how to react once a wrong answer has gone out. The causes of invented answers have their own guide on why chatbots make things up. Here the focus is responsibility and prevention.
Who is responsible for what the chatbot wrote?
The customer enters a relationship with your business, not with the model provider or the chat window. When an answer misleads someone about a price, a condition or a qualification, they have no way of knowing that this part of the page is less reliable than your terms. The safer assumption is that every sentence from the assistant will be read as your company speaking.
Towards the customer, it is hard to argue that the software made the mistake rather than an employee. A note saying chat answers are "for guidance only" does not replace control over what the assistant says. The best protection is fewer situations in which a wrong answer can arise.
Two decisions: Air Canada and OLG Hamm
In Moffatt v. Air Canada, the airline's chatbot told a customer he could apply for a bereavement fare after travelling. The page describing that fare said the opposite. Air Canada argued it could not be held liable for information provided by the chatbot. The tribunal called this a remarkable submission: the chatbot is part of the company website, and a customer should not have to double-check one part of the site against another.
Scroll the table sideways to see all columns.
| Case | What the chatbot said | Outcome |
|---|---|---|
| Moffatt v. Air Canada, Civil Resolution Tribunal (British Columbia), 14 February 2024 | A bereavement fare could be claimed retroactively, after travel | Negligent misrepresentation; the airline was ordered to pay CAD 812.02 in total |
| Verbraucherzentrale NRW v. Aesthetify GmbH, OLG Hamm, 12 May 2026, case 4 UKl 3/25 | The company doctors held specialist titles in plastic and aesthetic surgery they did not have | Chatbot replies count as the company's own commercial conduct; injunction granted, appeal to the BGH allowed |
In the Hamm case, the chatbot of a business offering aesthetic treatments gave false information about its doctors' specialisations. The court held the company responsible for those statements even if the chatbot had been fed only correct data, and found that the chatbot is not a "third party". The court allowed an appeal to the Federal Court of Justice, so this may not be the final word.
The full reasons in the Canadian case are published by the Civil Resolution Tribunal (2024 BCCRT 149). It is a small-claims decision from British Columbia, not a precedent for courts elsewhere.
The OLG Hamm judgment is summarised in a press release from the North Rhine-Westphalia judiciary. It concerns German unfair competition law (UWG) and an injunction sought by a consumer association; it does not decide damages claims by individual customers.
Questions where a mistake costs the most
Not every error has the same consequences. Wrong opening hours are annoying, but a wrong price, deadline or approved exception can lead a customer to a decision they would not otherwise make. List the topics where an answer from your assistant works like a commitment from your company.
- Prices and discountsAmounts, promotions, delivery costs and refunds, stated only as the current price list gives them.
- DeadlinesDelivery times and periods for withdrawal, complaints or cancellation. A deadline without a source is worse than pointing to your terms.
- ExceptionsRequests to bend the rules. A named person decides; the assistant only explains how to ask.
- Promises and qualificationsClaims about guarantees, results, completed actions or staff credentials. Every word must come from a verified source.
For each topic, note where the binding source lives and who keeps it up to date. If no such source exists, the assistant should not answer on that topic by itself. Route requests for discounts or exceptions straight to a person. At a medical practice, symptoms, medication and test results are topics the assistant should never answer by itself, as described in the guide on AI chat for clinics.
Answers only from an approved knowledge base
Your assistant should answer from a curated knowledge base, not from the general knowledge of the model or random pages. Curated means someone chose the sources, removed contradictions and is responsible for keeping them current. When the price list changes, the old version must leave the knowledge base the same day. In the Air Canada case, the chat answer contradicted the fare policy page.
Behaviour when information is missing matters just as much. An honest "I don't have that information" beats a plausible guess, as long as the customer gets a way forward. That way forward is a handover to a person with the customer's consent, without promising a quick reply when nobody on the team is working. The flow is described in the guide on handing an AI conversation over to a human.
In Meet AI, the assistant uses only the active knowledge base. When it finds no reliable passage, it does not call the model; it says the information is not available and offers a conversation with the team, and the handover happens only after the customer agrees. The default Copilot mode does not answer customers at all: it suggests a draft that an agent reviews and sends. This lowers the risk but neither guarantees error-free answers nor shifts responsibility away from you.
Private tests before the first customer conversation
Before the assistant talks to customers, ask it the high-risk questions and write down in advance what a correct answer would be. Add variants with a false premise, such as "since returns are possible after 60 days, how do I file one?", plus requests for exceptions. A correct result is sometimes a refusal or a handover. A full set of control questions is in the guide to testing a chatbot before launch.
- Every price in an answer is backed by the current price list.
- Deadlines match your terms and are never rounded or extended.
- A request for an exception ends with the name of the decision-maker, not with approval.
- A question outside the knowledge base produces a missing-information notice and an offer of contact.
- No answer suggests that an action has already been carried out.
Private tests in Meet AI show the result together with citations of the sources actually used and have no public effects. A test that produces an answer uses a generation from the shared quota, so prepare your question set in advance.
When a wrong answer has already reached a customer
Stop the source of the problem first, then deal with the customer and the records. The procedure below is a suggestion for a small team, not a legal requirement.
- Preserve the conversation.Keep the full transcript, date and answer. Delete nothing, even if it is embarrassing.
- Withdraw or fix the source.Delete or fix the faulty knowledge source, and answer that topic yourself until the correction is live.
- Check whether the customer acted.Find out whether the customer bought, booked or cancelled because of the answer.
- Hand the decision to an authorised person.Someone with the authority decides whether to honour what the chatbot said, and a lawyer joins if there is a dispute.
- Correct the information and record the lesson.Send a plain correction and add the case to your test set so the error does not return after the next knowledge update.
In Meet AI, deleting a source or unpublishing a Help Center article removes its passages from active knowledge immediately. How to word the correction for the customer is covered in the guide on correcting a wrong answer. If you want to find out why the assistant got it wrong, work through the causes of invented answers.
Questions about chatbot responsibility
Does labelling the chat as AI remove the company's responsibility? Do not count on it. A label tells customers who they are talking to, but it does not make a false answer true, and both cases above looked at what the answer said. The rules for labelling itself are covered in the guide on labelling an AI chatbot.
Has a court in my country ruled on this? Beyond the two cases above, this article does not predict how local courts would decide. A reasonable working assumption is to treat assistant answers the same way as answers from an employee. In an actual dispute, get legal advice.
