An AI chatbot gives wrong answers when it fills gaps from the model’s general knowledge instead of your approved content, when that content is outdated or contradicts itself, or when nothing tells it to admit it doesn’t know and hand over to a person. Fix it in this order: decide what the bot may answer, write the answers it must always get right, restrict it to those sources, let it say “I don’t know”, then test with a fixed set of real questions and review live chats. These steps reduce wrong answers but can’t remove them, and your business still answers for what the bot says.
Why does an AI chatbot give wrong answers?
AI chatbots are built on language models, and models sometimes produce text that sounds right and isn’t. This is called a hallucination. Anthropic’s guide to reducing hallucinations says even the most advanced models can generate text that is factually incorrect or inconsistent with the given context. OpenAI researchers argue in “Why Language Models Hallucinate” that models hallucinate partly because training and evaluation reward guessing over acknowledging uncertainty. In a business chatbot the result is a confident answer about a price, a policy or an opening time that nobody approved.
The model is rarely the only cause. Most wrong answers in a business chatbot trace back to what the bot was given, what it was told to do when it doesn’t know, and how it was tested. Match the symptom to the likely cause:
| What you see | Likely cause | Fix |
|---|---|---|
| It invents a policy, price or feature that doesn’t exist | It answers from the model’s general knowledge because no approved source limits it | Restrict it to your documents and let it say “I don’t know” (steps 3–4) |
| It quotes an old policy or price | The content it reads is outdated, or two versions contradict each other | One current source for each fact, old pages removed (step 5) |
| The right document, the wrong passage | The search step that fetches passages returns the wrong part of a page | Rewrite long pages into clear questions and answers, then re-test (steps 5–6) |
| It answers refund, medical or legal questions it should pass on | No rule says which topics need a person | Write the hand-off rules (steps 2 and 4) |
| It is right in testing but wrong with customers | The test used a handful of easy questions | Build the test set from real customer questions (step 6) |
| It gives different answers to the same question | Nothing checks consistency | Ask the same question several times in testing and compare (step 6) |
Why wrong answers are your problem: the Air Canada case
In February 2024 a British Columbia tribunal decided Moffatt v. Air Canada (2024 BCCRT 149). A customer asked the airline’s website chatbot about bereavement fares and was told he could buy a regular ticket and request a partial refund within 90 days. The airline’s staff later told him refunds could not be claimed afterwards. Air Canada argued that it could not be held liable for what its chatbot said and described the chatbot as a separate legal entity. The tribunal rejected that, found that the airline did not take reasonable care to ensure its chatbot was accurate, and ordered a partial refund of CAD 650.88 plus damages for interest and tribunal fees, according to the law firm Manatt’s summary of the decision.
It was a decision of one tribunal in one province, so treat it as an example, not as legal advice for your country. The tribunal questioned why a customer should have to double-check the chatbot against other pages of the website. Check what your bot says the way you would check a page on your site.
Before you change anything
- Keep every wrong answer. Screenshot or export the full conversation, with the date.
- Don’t swap the AI model first. A different model can still guess, and changing it first makes the real cause harder to see.
- Pull the bot off costly topics until it is fixed. Send refunds, medical questions, legal questions and anything about money owed to a person.
- Find out who owns the content the bot reads: policies, prices, FAQs and help pages. Someone has to keep them current.
How to fix wrong answers, step by step
Step 1: Sort the wrong answers
List each wrong answer in a table: the question, what the bot said, what the correct answer is and where that answer lives on your site. Tag each one with a cause from the table above. This list becomes your test set in step 6.
Step 2: Decide what the bot may answer
Write down the topics it handles, such as opening hours, shipping or how booking works, and the topics it must pass to a person, such as refunds, complaints, medical or legal questions. Then write the exact answers to the questions it must always get right. This is the knowledge step in the way we build chatbots at Auroxa: we load the website, documents and FAQs and write the answers the bot must always get right.
Step 3: Ground it in your content
Tell the bot to answer only from the documents you provide, not from general knowledge. Anthropic’s guide lists this as external knowledge restriction. It also suggests asking for a supporting quote for each claim and removing any claim the bot can’t support. For policies and prices, show a link to the source next to the answer.
Step 4: Let it say “I don’t know” and hand over
Anthropic’s guide says explicitly allowing the model to say it doesn’t know can drastically reduce false information. The OpenAI paper above explains why this has to be deliberate: models tend to guess unless the setup makes admitting uncertainty acceptable. Write the hand-off as a rule. When the answer isn’t in the sources, the bot says so, offers a person by email, phone or a booking link, and records the question. Test the rule, because one sentence in the instructions won’t always be followed, and a bot that refuses everything is no use either.
Step 5: Clean the knowledge
- Remove outdated pages and duplicates, or mark them clearly as old.
- Keep one source for each fact, with a date and an owner, such as one refund policy and one price list.
- Add answers for the questions from step 1 that the documents don’t cover.
- Load the content again after every change, then re-test.
Step 6: Test with a fixed set of real questions
Build a list from the questions customers actually ask, plus a few that your documents can’t answer. For those few, the right result is “I don’t know” or a hand-off, not a guess. Run the list before and after every change and count how many answers are right. Anthropic also suggests running the same prompt several times and comparing the outputs, because inconsistent answers can signal a hallucination.
Step 7: Review real conversations
Read a sample of live chats every week at first. Add each wrong answer to the test set and fix its cause. Our chatbot service includes monthly reviews and care after launch.
Step 8: Keep people in control of actions
If the bot can take actions, such as booking a call or creating a support ticket, decide which ones it may do on its own and which need a person’s approval. A wrong answer is a problem. A wrong action costs more.
What these fixes can’t do
Anthropic’s guide is clear that its techniques significantly reduce hallucinations but don’t eliminate them, and it advises validating critical information. Plan for the errors that remain: link the official policy next to answers about money, keep a person one step away and review chats on a schedule.
We build AI chatbots with these checks from the start. If you already have a bot that gets things wrong, get in touch and bring ten of its wrong answers.
Frequently asked questions
What is an AI hallucination?
A confident answer that is wrong or not supported by the information the model was given, such as an invented policy, price or feature.
Can an AI chatbot be stopped from giving wrong answers completely?
No. Anthropic’s guide says its techniques significantly reduce hallucinations but don’t eliminate them. Aim for a low, measured error rate, a hand-off to a person and regular reviews.
Is a business responsible for what its chatbot says?
In Moffatt v. Air Canada, a British Columbia tribunal held the airline responsible for its chatbot’s wrong information. That decision applies in one province, so ask a lawyer about your own country. Check the bot’s answers the way you would check your website.
How do I test a chatbot before launch?
List real customer questions plus some your documents can’t answer. Run them, mark each answer right, wrong or hand-off, and repeat after every change. For a question that isn’t in your documents, “I don’t know” counts as a correct answer.