Hallucination
When an AI states something false, invented, or unsupported as if it were fact: a fake citation, a wrong date, a feature that does not exist. It happens because the model produces the most plausible next words, not checked ones. Plausible and true overlap most of the time, not always. Every name, number, and link needs checking.
In plain English
The model is a confident intern who never says “I’m not sure.” Ask for a source and it will give you one. Sometimes that source doesn’t exist. It isn’t lying. It’s producing the most likely-sounding answer, and likely-sounding is not the same as true.1
In the real world
- A citation to a paper that was never written.
- A “feature” of your software that the model invented from the name.
- A biography with the right shape and the wrong facts.
When to use it
Name it when you see it, so the team stops saying “the AI lied.” It didn’t. It guessed. The fix is different.
The two-minute check:
- Ask for the exact quote and where it came from.
- Search the name, title, or number yourself.
- If you can’t find it in two minutes, it doesn’t go in.
When not to use it
Don’t use it as a reason to avoid AI. Use it as a reason to check. Anthropic’s own advice: let the model say “I don’t know,” ground it in a document, and ask for quotes you can verify.2
Good, bad, ugly
- Good: once you know it happens, you build checking into the workflow.
- Bad: the false parts read exactly like the true parts.
- Ugly: a hallucinated fact in a client deliverable, found by the client.
Key facts
- Looks like
- Invented quotes, sources, statistics, URLs, and product features
- Why it happens
- The model produces likely text, not checked text
- Reduce it with
- Let it say "I don't know," give it the source document, ask for quotes you can verify
Common questions
Why does the AI make things up?
It's built to produce the most plausible next words. Plausible and true overlap most of the time, not all of the time. When they don't, you get a hallucination.
How do I catch one?
Ask for the exact quote or page number and check it. Anthropic's guidance includes letting the model say it doesn't know and having it cite the text it relied on.
Will this ever go away?
It's shrinking, not gone. Treat every fact, name, number, and link as unverified until you've looked.
References
- AI hallucinations, explainedIBMView source
- Reduce hallucinationsAnthropic, Claude Platform DocsView source
- Hallucination (artificial intelligence)WikipediaView source