AI capability evaluation is a mature field. Harm evaluation is not, and the asymmetry is structural rather than accidental. The researchers building model benchmarks are answering questions the labs and funders commissioned: can the model pass the bar exam, write working code, reason across long contexts? These are answerable questions with clean metrics. The question of what AI-generated content is doing to the people who encounter it — to their beliefs, their safety, their sense of what is real — has in active institutional use.…