Google DeepMind is testing double-blind evaluations to remove human bias from AI benchmarking. This method hides both the model identity and the evaluator's identity during testing. It prevents brand prestige from inflating scores.