Hallucination

A hallucination is a model stating something false with full confidence and no signal that it is guessing. The behavior follows from how generation works: a language model produces the most plausible next words, and where it has no grounding for a fact, the most plausible-sounding answer and the true answer differ. The gap gets filled fluently, with no flag.

Prompting cannot remove the behavior, because plausible-next-words is the whole mechanism. Three mitigations work, all structural: ground answers in retrieved sources so a claim traces back to something real, let the system refuse when it has no support instead of inventing one, and test for the confident wrong answer, the most expensive kind.