AI hallucination rates are now benchmark-dependent. A model can look solid yet...

https://wiki-club.win/index.php/Is_HaluEval_Broken_if_a_Length_Rule_Gets_93.3%25_Accuracy%3F

AI hallucination rates are now benchmark-dependent. A model can look solid yet fail at 30.2% on the HalluHard test. Whether you use Vectara HHEM or AA-Omniscience, your choice of metric defines your risk