Spotting a hallucination

What a hallucination is (and why it happens)

A hallucination is when AI tells you something that is simply not true, but says it with total confidence. It is not lying — it genuinely does not know whether it is right. AI works by guessing what words usually come next, and most of the time that lands on the truth. But when it does not, you get a made-up quote, an invented study, or a number that is just wrong — delivered exactly as confidently as the true stuff. That is the catch: you cannot tell by how sure it sounds. You have to bring the doubt yourself, and the prompts below do most of that for you.

What one looks like

Say you ask: 'What did the CEO of a company say about working from home in 2023?' A hallucination looks like this — and reads completely convincingly:

In a March 2023 interview with the Financial Times, she said: "Remote work is here to stay, and any company that fights it will lose its best people."

Make the AI admit what it is guessing

Paste this at the end of any answer you are not sure about. It makes the AI separate what it actually knows from what it is guessing at.

For each thing you just told me, how sure are you from 1 to 5, and why? Point out anything you are guessing at rather than actually knowing, and list any names, dates, numbers, or sources I should double-check before I trust them.

Ask it to prove a source exists

The most common made-up thing is a source — a study or a link that looks real but is not. Ask the AI to back it up, and watch what happens when it cannot.

Where exactly does that come from? Give me the title, who wrote it, and where I can find it. If you are not sure it really exists, tell me that instead of guessing.

The quick check

You do not need to check everything. Just build a habit for the things that matter — anything to do with money, health, the law, or a decision you cannot take back.

1. 1Ask for the source (prompt above), then open it yourself. Never trust the AI's summary of something you have not read.
2. 2Search the exact claim on your own. If it is true, you will usually find it somewhere else too.
3. 3Ask again in a new chat, worded differently. If the answer stays the same, it is probably solid. If it changes, be careful.
4. 4If it matters, treat the AI's answer as a first draft — never the final word.

Stop it before it starts

You can also tell the AI, up front, to admit when it is unsure — so you are not chasing it later. Add this to your custom instructions, or paste it at the top of a chat.

When you are not sure about something, say so instead of guessing. Never make up quotes, studies, statistics, or links. If you do not know, just say you do not know.

The one rule to remember

AI is neither magic nor a menace. It is a tool that is brilliant at some things and quietly unreliable at others, and the whole skill is telling which is which. Use it freely to draft, reword, brainstorm, and get unstuck — anywhere a wrong detail is easy to spot and cheap to fix. Do your own checking anywhere a confident wrong answer would actually cost you. That is not being paranoid; it is just knowing where the edge is and staying on the right side of it.