Why does an AI chatbot make things up, and how to fix it (2026)

Every so often an AI chatbot answers confidently, and wrongly. People call it a "hallucination", and it's the number one worry before deploying one. The good news is that when a bot answers from your own content, the way Breezaro does, it's largely preventable, because you control what it draws on. Here's why it happens, and three fixes that work.

Why it happens

A language model's instinct is to produce a fluent answer even when it doesn't actually know. Answering from your content curbs that (see how an AI chatbot works), but the bot can still slip when:

• the answer isn't in your content, so it fills the gap with a plausible guess,
• your content is vague or contradictory, so it stitches together the wrong pieces,
• the question is ambiguous and it guesses at your intent.

Almost always, an invented answer is a content problem rather than a technology problem, which means you can fix it.

Tighten the prompt

The bot's instructions (the system prompt) set its behaviour, and a few lines make a big difference:

• Tell it to answer only from the provided content, and to say it doesn't know rather than guess.
• Set the tone and scope ("you are the support assistant for our shop; answer about our products, orders and policies").
• Tell it what to do at the edge, such as offering to hand off to a human or pointing to your contact channel.

You can edit this any time in the dashboard, and changes apply immediately. (More in customize your chatbot.)

A word of caution. The prompt cuts both ways. Our default is already tuned for grounded, honest answers, so heavy edits can backfire. A bot pushed to be maximally "helpful" may start promising things it can't do, like arranging a meeting, sending an email on your behalf, or committing to a deadline. If answers drift after an edit, restore our default with the Reset prompt button, then make smaller, targeted changes to steer it.

Fill the knowledge gap

The most common cause is simple. The answer isn't in the bot's content, so it improvises. The fix is just as direct. You add the missing answer. Read the conversations where the bot struggled, add that content, and next time it answers correctly from your material. This is the core habit of training your chatbot well. In the meantime, an operator can take over.

Structure your content well

How your content is written changes how reliably the bot finds the right passage:

• One topic per section, with a clear heading. A focused passage retrieves cleanly, while a wall of mixed topics doesn't.
• Be explicit and specific. "Returns within 14 days of delivery" beats "flexible returns", because the bot can only be as precise as your source.
• Remove stale and contradictory content. Two different return windows in two documents is a recipe for a wrong answer. Keep one source of truth, and keep it current.
• Prefer clean documents. A tidy FAQ or policy page beats a scan of a scan of a PDF.

Putting it together

The three fixes compound. Clear instructions, complete content, and clean structure add up to a bot that answers from your facts and defers when it can't. You won't reach zero mistakes (no support channel does), but you can get to "rare and fixable", with a human ready for the rest.

FAQ

What is an AI chatbot "hallucination"? It's when the bot gives a confident but incorrect answer, usually because the real answer isn't in its content, so it guesses.

Can I stop it completely? Not entirely, but a bot that answers from good content, with a tight prompt and human takeover, keeps it rare and correctable.

Why does it make things up instead of saying "I don't know"? Language models default to answering. You steer it toward "I don't know" through the prompt, and toward the right answer by filling content gaps.

The bot gave a wrong price. What now? Find the source with the wrong or missing price, correct it, and sync it again. Wrong answers almost always trace back to the content.

Invented answers feel like a flaw in the technology, but they're usually a signal about your content. Get those three things right, and the bot becomes something you can trust on the front line.

Related: How does an AI chatbot work? · How to train your AI chatbot.