An analysis of 1,629 benchmark results from 180 AI models revealed that AI labs tend to report scores that are, on average, 6.4% higher than those reported by independent testers. This discrepancy was observed across 103 different benchmarks. The project, named whatmodel.sh, has made its dataset and code publicly available under open-source licenses to promote transparency in AI model performance reporting. AI
IMPACT Highlights potential bias in AI model performance reporting, urging for greater transparency.
RANK_REASON Analysis of benchmark scores comparing AI labs and independent testers. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →