A hallucination is output from a generative AI model that sounds convincing yet is factually wrong or entirely made up. Typical cases are invented references, court decisions that never existed, wrong figures or software functions that were never built. The cause lies in the method: a language model produces the most probable continuation of a text and does not check facts while doing so.
Hallucinations can be reduced, yet never fully ruled out. Proven measures are grounding in verified sources, a requirement to supply evidence and expert review before results leave the organization or feed into decisions. These rules belong in the AI Usage Policy. The threat T13 Hallucination / faulty output describes examples and controls.