A hallucination is a model-generated statement presented as fact that is false or unsupported, produced because the model predicts fluent text rather than verifying truth.
How it works
A language model learns patterns from training text and, at inference time, emits the most likely continuation of a prompt without consulting a source of truth. When the prompt asks for something beyond its training, the same machinery that writes smooth prose fills the gap with plausible invention: fake citations, wrong numbers, confident nonsense. Retrieval grounding, which pairs generation with passages pulled from real documents, cuts the rate sharply, as do evals that score factuality on benchmark suites instead of style alone. Lowering temperature steadies wording but cannot fix missing knowledge, so anything unverified stays suspect until a source confirms it.
Builders treat every ungrounded answer as guilty until proven innocent: grounding responses in retrieved documents, asking for quotes beside claims, and routing high-stakes facts through checks before they reach readers.