In 2026, "accuracy" is just marketing noise. Hallucination rates shift wildly...

https://www.tumblr.com/gladlyradiantsphinx/816924113252859904/stanford-ai-index-why-documented-ai-incidents

In 2026, "accuracy" is just marketing noise. Hallucination rates shift wildly depending on your chosen benchmark. For example, the HalluHard suite captures a 30.2% failure rate in complex reasoning that simpler tests miss entirely