Hallucination
What is a hallucination?
A hallucination is content an LLM makes up: a detail, a quote, or a fact that isn't in its source material or isn't true. LLMs are probabilistic. They generate text that is plausible given the input, not text that has been verified against a source. A model summarizing a transcript can state a job title that was never mentioned or invent a quote that sounds like the participant.
Why do hallucinations matter for product teams?
If you use AI to summarize customer interviews, synthesize customer feedback, or write up meeting notes, hallucinations undermine the point. The summary may be telling you something customers never said.
When Teresa Torres used ChatGPT to turn seventeen interview transcripts into stories, her error analysis surfaced two errors. It hallucinated details, like a participant's job title. And it hallucinated quotes, merging two statements into a new meaning that wasn't in the transcript. Those errors became her first two evals: a fact checker and a hallucination guard.
How do you catch hallucinations with evals?
Start with the cheapest measurement that works. For fabricated quotes, a code assertion is enough: for every quote in the output, search for the exact string in the source. A match returns true; no match returns false. This is Teresa's hallucination guard. It requires no judgment, so it's fast and virtually free.
Fabricated facts are harder. A model can invent a claim using different words, so there's no string to search for. That calls for an LLM-as-a-Judge fact checker: the judge gets one claim and the transcript and returns true if the claim is grounded in the transcript, false otherwise. Together the two evals give you an error rate and tell you which quotes and facts to fix by hand.
Related terms:
Last Updated: September 1, 2026