To maximize the benefits to humanity, decisions to deploy AI solutions must be based on comparing AI with the average performance of unaided humans, not an unrealistic comparison with the single best human or with unachievable perfection.