A hallucination is output from an AI model that sounds confident and plausible but is false or unsupported — an invented citation, a fabricated figure, a made-up policy detail, a quoted source that doesn't exist. The term captures the unsettling quality of the failure: the model isn't hedging or erroring out, it is asserting something untrue with the same fluency it uses for correct answers.
Hallucinations happen because of what a language model fundamentally does. It predicts plausible continuations of text based on patterns in its training data; it has no built-in mechanism that checks claims against reality. When the model lacks the specific knowledge, when the prompt pushes it toward an answer it can't actually support, or when the right information simply isn't in its training, it fills the gap with something that fits the pattern. The output is statistically likely text, not verified fact — and likely text can be wrong.
In production, hallucination is the central reliability risk of using language models, and most of the architecture of a serious AI system exists to control it. Retrieval-augmented generation grounds answers in retrieved source documents so the model summarises real material instead of recalling from memory, and lets you cite the source. Prompts instruct the model to answer only from supplied context and to say when it doesn't know. Guardrails and evaluations test specifically for unsupported claims. None of these eliminate hallucination entirely, but together they reduce it from a constant hazard to a managed, measurable one.
Hallucination matters because a confidently wrong answer can be worse than no answer — it erodes trust, and in regulated or high-stakes settings it creates real liability. The practical lesson for anyone deploying AI is that the question is never just "is the model smart" but "how is this system designed so that wrong answers are rare, caught, and traceable." Treating hallucination as an engineering problem to be measured and bounded — rather than a flaw to be wished away — is what separates a demo from a system you can put in front of customers.