A hallucination is output from a language model that reads as fluent and confident but is false, or isn’t supported by the material the model was given. A model that cites a paper that doesn’t exist, attributes a quote to the wrong person, or adds a detail to a summary that the original never mentioned is hallucinating.
The term predates chatbots. A 2022 survey by Ji and colleagues traces it through machine translation, summarization and dialogue research, where it meant generated text that is nonsensical or unfaithful to the source. It borrows the psychological sense of a perception that feels real with nothing behind it.
Two kinds
The survey sorts hallucinations by their relationship to the source:
- Intrinsic: the output contradicts the source. A summary says a vaccine was approved in 2021 when the article says 2019.
- Extrinsic: the output adds something the source can neither confirm nor refute. It might even be true, but nothing in the input supports it.
In open-ended chat there’s often no source document at all, so in everyday use “hallucination” has come to mean any confident factual error.
Why models do it
A language model is trained to produce likely text, not to check facts. Kalai and colleagues argued in 2025 that hallucinations are ordinary statistical errors: if true statements and plausible false ones can’t be told apart from the training data, some mistakes are unavoidable. They add a second cause. Most benchmarks give no credit for “I don’t know,” so a model that guesses scores better than one that admits uncertainty, and training follows the scores. Their paper argues the fix is partly in how evaluations are graded.
That’s why the diagram’s two columns look alike on the surface. Fluency and confidence come from the same machinery whether the content is right or wrong, so tone tells you nothing. What’s missing is calibration: confidence that tracks how often the model is actually right.
What reduces it
- Grounding answers in supplied text, most often through retrieval-augmented generation, and asking for citations that can be checked.
- Tools that look facts up or run calculations instead of recalling them.
- Room to abstain: prompts, training and scoring that reward “I’m not sure.”
- Verification: tests for code, source checks for claims, a person for anything high-stakes.
None of these brings the rate to zero, so unchecked model output is best treated as a draft.