Unverified50% confidenceBenchmarkExact time
The gpt-4o-mini judge mislabeled as high as 77.1% of human cases on the enriched set that should have been judged as faithful
1
Sources
50%
Confidence
Long-term
Relevance
9/28/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedInternal research from companies such as Microsoft and Google indicates that even GPT-4-level models may have factual error rates as high as 15%–30% in vertical domains such as law, finance, and healthcare.72% similarUnverifiedGPT-4 as a judge can achieve over 80% agreement with human evaluations, but exhibits systematic issues such as position bias, verbosity bias, and self-preference.69% similarUnverifiedIn an unknown-rules mini-game test designed specifically for models, GPT-6 scored 99 points, approaching a perfect score, while the highest human score was only in the 30s and other models scored around 40.65% similarUnverifiedIn 96% of levels, GPT-6 required fewer attempts to find the answer than the average person.65% similarUnverified研究表明人类评估者在判断LLM输出的事实准确性时，正确率往往只有50%-70%64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/955610API
curl https://kongchang.com/api/v1/knowledge/claims/955610MCP
get_claim(id=955610)