Skip to content
Grok Academy
All resources
Tutorial · 6 min

Why do models hallucinate?

What researchers mean by hallucination, when it shows up, and how to catch it in a Grok thread.

Hallucination, in this context, is fluent content that does not match the world or the sources you provided. It is not a bug that will be patched away entirely. It is what next-token prediction does when the prefix does not constrain the answer hard enough.

  • You asked for a citation, a quote, or a number that was not in context.
  • You asked a very specific question in a sparse domain.
  • You rewarded confidence ('be definitive') instead of calibrated uncertainty.

Tactics: require sources and open them; ask for [UNKNOWN] on gaps; split 'what do we know' from 'what do you infer'; prefer tools for anything you will bet money on.