Hallucination rate
Also known as: hallucination
Hallucination rate is the fraction of a model’s outputs that are false or unsupported — a concrete, measurable form of hallucination.
You can measure it:
- Have a judge check each claim against a source.
- Count unsupported statements.
- Drive the rate down with grounding, citations, and “say ‘I don’t know’”.
Set expectations: no model has a zero hallucination rate. The goal is to *measure* it, keep it low, and make the model cite its sources so you can verify.