Hallucination rates depend on the benchmark. Vectara HHEM and AA-Omniscience...

https://holdensinsightfulblogs.wordpress.com/2026/05/18/why-did-gpt-4o-accuracy-drop-to-64-4-when-the-user-believed-something-false/

Hallucination rates depend on the benchmark. Vectara HHEM and AA-Omniscience reveal different failure modes. With $67.4B lost to bad data, stop trusting vendor averages. Demand testing that mirrors your unique, real-world operational risks