When a large language model produces fluent, confident text that is false or unsupported, because it predicts probable wording rather than drawing on verified facts.
Hallucination is when a large language model produces fluent, confident text that is false or unsupported by any source. The model generates the most probable continuation of a prompt from statistical patterns in its training, not from a store of verified facts, so a wrong answer reads exactly as convincingly as a right one.
It is the default failure mode whenever the model lacks a fact, holds a stale or garbled version of it, or is pushed to answer beyond what it knows. Common forms are invented statistics and dates, fabricated citations and URLs that never existed, and confident misattribution of a quote or a product feature to the wrong brand.
Grounding and retrieval-augmented generation reduce it by giving the model real sources to answer from and to cite, though neither removes it: a model can still misread a retrieved page or ignore it. For AI visibility this cuts both ways. A hallucinated fact about a brand can spread through AI answers, and a brand absent from the sources a model can retrieve is more likely to be described from a hallucinated prior than from anything current.
