How to stop an AI chatbot from making up answers

An AI chatbot that invents an answer is worse than no chatbot. A visitor who reads a confident, wrong refund window acts on it, and your team cleans up after it. The good news: made-up answers are not random bad luck. They come from how the bot is set up, and you can control most of it.

Why chatbots make things up

A language model is built to produce a fluent answer to any question. When it has no reliable source, it fills the gap with what sounds plausible. Three setups make this more likely:

• The bot answers from general knowledge. It knows how refunds usually work, not how yours work.
• The bot has your documents but is allowed to go beyond them. When retrieval finds nothing useful, it answers anyway.
• Your documents disagree. An old policy and a new one both exist, and the bot picks the wrong one.

Five controls that stop it

1. Answer only from your documents

The bot should search your content first and answer only from the passages it finds. This is called retrieval-augmented generation (RAG). In Paperlink AI, the bot finds the best passages for each question and the answer model is told to answer only from them and to cite each fact.

2. Make "I don't know" a normal answer

If no passage is relevant, the right answer is a refusal, not a guess. Paperlink AI tells the visitor it does not know and shows your contact details instead. Replies the bot can't answer are free, so a refusal never costs you a reply.

3. Let your team see the source of every answer

You can't trust what you can't check. Paperlink AI shows your team the document and page behind each answer in the dashboard. When an answer is wrong, the source tells you whether the bot misread a page or the page itself is wrong.

4. Decide which document wins when two disagree

Remove outdated pages from the sources the bot reads. When you must keep both versions, mark the newer one. In Paperlink AI, you mark an uploaded document as superseding an older one. If the newer document also has a future effective date, it takes over on that date.

5. Review the questions the bot could not answer

Every refusal is a gap in your content. Paperlink AI collects them in the Unanswered report. Write the missing page, add it, and the next visitor gets an answer.

How to test a chatbot for made-up answers

Before you trust any bot, run this test with your own documents:

1. Write 20 questions your documents answer, and note the correct answer for each.
2. Write 10 questions your documents do not answer, but that sound like they should. For example, ask about a product you don't sell or a policy you don't have.
3. Ask all 30. Score each answer as correct, wrong, or correctly refused.
4. Any confident answer to one of the 10 trap questions is a made-up answer.

We run this kind of exam on Paperlink AI. In our internal evaluation (September 2026), the bot answered 93.8% of the held-back answerable questions correctly and refused 100% of the questions it should have refused.

Summary