Hallucination

What is Hallucination

Hallucination is when a model states something false or entirely invented with the same confident tone it uses for verified fact: an API method that does not exist, a citation to a paper that was never written, a statistic that sounds plausible but was never measured. The danger is not the error itself, it is that nothing in the model's tone signals the difference between something it knows and something it is filling in.

It happens most where the model has no real source to check itself against: asked to recall a specific detail it never actually saw, asked to summarize a document it was not actually given, asked for a number it has no way to have measured. This glossary's own hero stat for Skill was built specifically to avoid this: it reports a number read from a real file on disk, not a plausible-sounding estimate.

Hallucination is not reliably fixed by asking the model to be more careful, that changes tone, not accuracy. What actually works is grounding: giving the model a real source to read from and check against, and verifying claims against that source before they ship, rather than trusting a confident-sounding sentence on its own.