AI Glossary
Hallucination
model hallucination, AI confabulation
A hallucination is when a language model produces an answer that sounds credible but does not match the facts or the sources. It stems from how the model works, not from a malfunction.
- The model predicts likely text; it does not check whether it is true.
- It is most dangerous when the answer sounds confident and expert.
- You limit it by grounding answers in sources, by verification, and by evaluation.
A hallucination arises because a language model generates the most probable sequence of words rather than verifying whether the content is true. As a result, it can present a non-existent source, date, or number just as fluently as a correct fact, which makes the error hard to spot.
In practice you limit the risk of hallucination by grounding answers in specific documents (for example, through RAG), by adding verification, and by measuring the scale of the problem in model evaluation. Hallucination cannot be fully eliminated today, so in high-risk uses a person is kept in place to review the output.
Related terms
In guides