AI hallucination

An AI hallucination is a response that is fluent and confident but not supported by the model's sources or by fact, produced because the system is predicting plausible text rather than looking up a verified answer.

Inside a company the damage is different from the public examples. The invented output is rarely absurd; it is a plausible policy, a plausible threshold, a plausible named owner. It reads exactly like the real thing, so it gets acted on.

Grounding the model in internal documents reduces the rate but does not remove it, because retrieval can return nothing relevant and the model will still answer. The failure is not that the model lacks the facts, it is that nothing forces it to admit that.

The reliable control is refusal. A system that says it does not have enough covered material to answer, and names what is missing, is more useful than one that is right most of the time and gives no signal when it is not.