“According to researcher Melanie Mitchell, significant problems in AI evaluation include data contamination, where benchmark data appears in the training set.”