There's a lot of talk about regulating AI, but do regulators know the technology well enough? In our new paper, we survey major reg efforts & find they rely on benchmarking, which we know to be problematic. How did this happen & what can we do about it? arxiv.org/pdf/2501.15693