According to Ethan Mollick on X (Twitter), many AI benchmarks rely on smaller, cheaper LLMs as judges, but new research shows weaker judges cannot reliably evaluate stronger models; benchmarks should be viewed as a triplet of dataset, model, and judge, with judges becoming the saturated bottleneck (as reported by Ethan Mollick’s post on Feb 22, 2026). According to Mollick’s summary of the paper, evaluation quality degrades when judge capability lags behind the system under test, implying systematic bias and under-reporting of true model performance. As reported by Mollick, this creates business risk for AI product teams that optimize to flawed scores and highlights an opportunity for vendors offering stronger or calibrated judges, human-in-the-loop adjudication, and meta-evaluation frameworks. According to Mollick, the study urges benchmark designers to disclose judge-model specs, test judge consistency, and budget for higher-capacity evaluators when assessing frontier models.