AI evaluation methods face scrutiny over benchmark flaws and saturation

Recent research and testing have highlighted significant methodological flaws in how artificial intelligence models are evaluated. In a study by Everypixel, researchers attempted to benchmark AI image generation models by comparing them against real photographs. Using 863 blind judgments from 13 professionals, the study found that GPT Image 2 unexpectedly scored higher than real photographs, revealing that the benchmark's design was unreliable for measuring true visual quality.

Parallel academic research has addressed broader issues within AI evaluation, specifically the phenomenon of benchmark saturation. Studies indicate that as top-performing models improve, benchmarks often lose their ability to reliably distinguish between them. Furthermore, critics argue that the current benchmark culture may perpetuate structural harms and that the gamification of these evaluations presents a systemic challenge to the machine learning community rather than just a narrow methodological concern.