What is an AI hallucination?
An AI hallucination is output from a language model that sounds plausible but is false or unsupported, such as an invented citation, figure or policy.
In more detail
Models generate likely text, not verified facts, so they fill gaps with something that fits. The risk rises when the answer is not in the provided context, when the topic is obscure, or when the prompt pushes the model to always produce an answer. Reduce it by grounding (RAG with citations), telling the model to say when it lacks the information, constraining output to a schema and checking numbers in code instead of trusting the model's arithmetic. You cannot reduce it to zero. For high-stakes work, design so an error gets caught: show the source passage next to the answer and make a person confirm it.
Example
A contract assistant asked about a termination notice period answers '60 days' when the clause says 90. A grounded version quotes the clause with its page number, so a reviewer can verify the answer in seconds.