Major benchmarks measure what AI can do. None measure whether it does what you mean: the distance between what you ask an AI to do, and the unspoken assumptions about how you want the AI to do it. We propose a new metric: the Genie coefficient.