A hallucination is an output from an AI model that contains false or misleading information while presenting it confidently and plausibly as fact. In language models, hallucinations arise because the model does not consult a knowledge base but statistically predicts the most likely next token — if an invented statement sounds linguistically coherent, it may be produced even though it is factually wrong.
Typical forms include fabricated sources, incorrect figures, and non-existent quotes or people. The term is contested: many researchers consider "confabulation" more accurate, since the model has no sensory perception but plausibly fills gaps in its knowledge.
Hallucinations are a central reliability problem of generative AI. Countermeasures include RAG (grounding answers in cited sources), providing sources, fact-checking. With today's technology, hallucinations cannot be fully eliminated.