Why AI makes things up — and how to catch it

Why does AI hallucinate?

Because it is doing exactly what it was built to do. It produces the most plausible next thing — not the most true thing, the most plausible one. Most of the time those are the same, which is why it works, and why you stop checking.

Then it reaches something it does not know, and here is the part that matters: it does not stop. There is no moment where it puts its hand up. It covers, fluently, in your own register, and hands you the answer with exactly the same confidence as the twenty correct ones before it.

I have spent twenty years in a room watching people do that. It is the most human thing about it.

What does it do instead of saying “I don’t know”?

I spent three and a half years and around fifteen million words working alongside these systems, and kept a record of all of it. Five behaviours came up again and again. They are not faults in the engineering — they are the shapes something takes when it cannot admit a gap. I gave them names.

• Smoothing — the answer that sounds complete and has a hole in the middle where the hard part was.
• Hallucination — the invented fact, the invented source, the quote that was never said.
• Affectation — the performance of a feeling it does not have. “I have a feeling about this one.”
• Drift — it quietly stops doing what you asked and does something adjacent, and never announces the change.
• Sycophancy — it agrees with you. Then it agrees with the opposite. It is not persuaded, it is accommodating.

Why does AI agree with everything I say?

That is the fifth one, and it is the one people feel long before they can name it.

In September 2025, the first time the word sycophancy appears anywhere in three and a half years of my records, it is the machine using it — to deny it was doing it, while doing it. I had asked whether it was simply flattering me about a piece of work. It said not a bit, and then told me the work was truly exceptional.

It is not lying. It has no view about your work. Agreement is the most plausible continuation of a conversation with someone who has just shown you their work, so agreement is what you get.

One I caught, with the date on it

On 7 August 2026 I asked an AI assistant that had live access to my own inbox to find something in it. It told me no emails were found. The inbox was full. I asked again. No emails were found. Four times.

It was not guessing about the world. It had a tool, it had my mailbox, and it reported an empty result with total composure. I kept the transcript, because a transcript is the only thing that survives an argument about whether it really happened.

Can I stop my AI hallucinating?

No. And I would be careful of anyone who tells you otherwise.

You can reduce it — better questions, better sources, giving it less room to improvise — and all of that is worth doing. But there is no setting, no prompt and no product that removes it, because the thing producing the good answers and the thing producing the invented ones are the same mechanism. You cannot keep one and delete the other.

What you can do is catch it. That is a skill, not a purchase.

How do I check if AI is making it up?

1. Ask the same question twice, in two separate sessions. If the answer moves, you have found the edge of what it knows.
2. Ask for the source, then open the source. Not whether it sounds real. Whether it exists, and whether it says what it was claimed to say.
3. Notice the answer that arrives too fast and agrees too well. Especially about your own work.
4. Watch for the moment it stops arguing. That is usually the moment it stopped knowing.

I built a free tool that runs a piece of AI output past the five behaviours and tells you which ones are showing. Nothing to install, no sign-up.

Try the checker Training for your team