Models generate what's plausible, not what's verified — and when plausible and true diverge, you get a confident fabrication: a convincing citation to a paper that doesn't exist, a plausible library function that was never written. Fluency and truth come apart precisely because the training objective rewards the former.

Mitigations reduce it rather than cure it: retrieval (RAG) grounds answers in real documents, tool use lets claims be checked, and for code the compiler and tests are the honest referee — one reason coding is a favorable domain for AI. The reliable posture: treat unverified model claims as drafts.