Why does ChatGPT make things up, and how do I stop it?
Because it predicts plausible words, it does not look facts up. When it lacks the answer it produces something that sounds right, with total confidence. You cannot switch that off, only contain it: give it the source material, make it cite its sources, and check anything that matters before it ships.
Last updated 11 June 2026
It quoted you a regulation that does not exist, or a case study with a dead link, or a price your supplier has never charged. And it did it in the same calm, fluent tone it uses when it is right, which is the unsettling part. If it lied nervously you could spot it.
The behaviour has a name, hallucination, and it is not a bug being patched out next quarter. It comes from how these models work. ChatGPT does not contain a database of facts it consults. It generates text by predicting, word by word, what plausibly comes next, based on patterns in everything it read in training. Most of the time the most plausible continuation is also true. When the model has thin or no knowledge of your question, the most plausible continuation is a confident invention, because confident text is what it learned from. It is not lying; it has no idea it is wrong. There is no internal flag that distinguishes "I know this" from "this sounds like the kind of thing that is true".
Where this bites a small business
Not in brainstorming or drafting, where wrongness is cheap and you are reading every word anyway. It bites where output leaves the building unchecked: a chatbot quoting a price or policy to a customer, a proposal citing statistics nobody verified, advice on tax, employment law or health passed straight on. Those are also the cases where responsibility stays with you, not the model.
How to contain it
You cannot prompt it into honesty, but you can change the job from "remember" to "read", and that changes everything.
- Give it the source. Paste in the contract, the price list, the policy document, and say "answer only from this document, and say so if the answer is not in it". Grounded on real material, hallucination drops sharply. This is why pointing AI at your own documents is the pattern behind most serious business setups.
- Make it show its working. Ask for the source of each claim. Then check one or two. Fabricated citations collapse the moment you click them, which makes them a cheap tripwire.
- Use a model that searches. The paid tools can browse the web and link to what they found. A linked source you can open beats a remembered "fact" every time.
- Ask for uncertainty. "If you are not sure, say so" genuinely helps, but treat it as a seatbelt, not a guarantee.
- Keep a human on anything that ships. Customer-facing facts, numbers, legal and medical content get read by a person. Everything else can flow.
The mistake is using hallucination as a reason to dismiss the whole technology. The businesses getting value did something less dramatic: they stopped asking the model to know things and started handing it the things to read. I run AI systems all day, every day, and my rule has not changed in three years: no number, name or claim ships without a source I can open. Apply that one rule and hallucination stops being a danger and becomes what it really is, a known limitation you design around.
Answered by Dean Cookson, Founder and CEO at Operosus.