Hallucinations: why AI makes things up¶
Level: Beginner
An AI can make up a figure, a quote or a source with complete confidence. This tip describes the forms hallucinations take, the warning signs, and a verification method matched to what is at stake.
What a hallucination is¶
A hallucination is false information presented with the same confidence as accurate information. It is not a lie: the model has no intent. As explained in the tip What is an LLM?, it produces the most likely text, and a likely answer is not necessarily a true one.
The most common forms¶
- Figures and statistics: a precise percentage with no real origin.
- Quotes: a sentence attributed to someone who never said it.
- References: an article, a study, a court ruling or a law that does not exist, with a very credible title and date.
- Recent facts: information later than the model's knowledge, made up to fill the gap.
- Technical details: a software function, command or option that does not exist.
When the risk increases¶
- The more specialized or rare the question, the less reliable material the model has.
- The more precise the request ("give me three dated figures"), the more it is tempted to provide something, even if inaccurate.
- When you force an answer ("cite three sources"), it may fabricate some rather than admit it has none.
- When the topic touches a regulated field (legal, medical, financial), where a mistake is costly.
Warning signs¶
- Suspicious precision: a figure with two decimals, an exact date, with no explanation.
- A source given without a link or with a title that is too perfect.
- A very confident tone on a topic that is actually debated.
- An answer that changes when you ask the question again.
Three reflexes in practice¶
1. Ask for sources, then open them. Asking is useful, but proves nothing: a cited source can be made up. Check that it exists, then read the passage that is supposed to confirm the statement.
2. Provide the text instead of asking from memory. Paste or attach the document the AI must work on, after anonymizing it if needed (see the tip Data security and privacy). It then relies on what you give it rather than on what it thinks it knows.
3. Cross-check critical points. A figure, a date, a name, a legal reference: a single check against a reliable source avoids most errors.
Matching the check to the stakes¶
| Stakes | Example | Check |
|---|---|---|
| Low | Title idea, rephrasing | Quick proofread |
| Medium | Figure in an internal presentation | Check the source |
| High | Customer document, legal opinion, decision | Full verification and sign-off by a manager |
A step-by-step example¶
You ask the assistant for a market figure for a presentation. It answers with a percentage and the name of a study.
- You look for the study: it does not exist under that title.
- You remove the figure or replace it with sourced data.
- You note that, for this type of request, you will now provide the source yourself and ask the AI to stick to it.
The responsibility stays yours¶
The AI offers a working draft; you are the one who signs off on the result. Proofreading, checking and validating before any use is part of the method, not excessive caution.
Hands-on exercise (5 minutes)¶
- Ask a tool to cite three sources on a topic of your choice.
- Check whether each one really exists.
- Check whether it really says what the AI claims.
- Note how many are accurate.