AI hallucination: why your AI makes things up, and how to cut it down

AI hallucination is when an AI model gives a confident answer that isn’t true. It happens when the model doesn’t have the facts in front of it, sees a document in pieces, is asked for something that isn’t there, or has no example of a good answer. Giving it sources, a way to say “not found” and a worked example cuts most of it.

Hallucination is the biggest risk in putting AI in front of customers or staff. A wrong answer in the same confident tone as a right one is hard to spot, and people act on it.

Why AI models hallucinate

An AI model writes what sounds right. It doesn’t check facts. When it doesn’t know the answer, it still writes something that sounds right, in the same confident tone.

So the fix isn’t asking it to be careful. It’s giving it the facts, and a way out when the facts aren’t there.

It didn’t have the facts in front of it

A model answering from memory fills gaps with guesses. Give it the facts instead: search your documents for the passages that answer the question, and have it answer only from those. This is called RAG. See how to get an AI chatbot that answers from your company’s documents.

It saw the document in pieces

Long documents are often read in parts. A question about the whole document then gets answered from one part, and the AI fills the rest with something that sounds right.

When the document fits, show the model the whole document. When it doesn’t, search for the parts that answer the question, so the right ones are the ones it reads.

It was asked for something that isn’t there

Ask an AI model for a contract’s end date and it will give you a date, even when the contract has none. The instructions need a way out: “If the document doesn’t say, answer ‘not found’.”

It had no example of a good answer

A worked example in the instructions shows the model what a good answer looks like, from a real document: what to pull out, in what form, and what to do when something is missing. One clear example often does more than another page of rules.

It wasn’t asked to show where

Ask for the exact words behind each answer. A made-up answer has no passage to quote, so it’s easy to spot, by a person or by a check in the code.

Measure it

Hallucination can be reduced, not removed. Run the AI on questions or documents where you already know the right answers, and count how often it gets them right. Include questions it can’t answer, to check it says “not found”. Then every change to the instructions or the model can be scored, and kept only if it helps. See how to know if your AI feature gives the right answers.

FAQ

Is a bigger model the fix? Not usually the first one. Try the fixes above before paying for a larger model.

Can hallucination be stopped completely? No. It can be made rare, easy to spot with sources, and measured, so you know how often it happens.

Does this apply to scanned documents? Yes, after the text is read out of the scan. Check that step first: a misread scan gives the AI the wrong text to begin with.