Benchmarks Miss Most American Work

A new mapping project suggests that current frontier AI benchmarks fail to cover roughly 90% of U.S. working hours, implying that benchmark performance overstates how much real jobs are automatable today. That gap means policy, business planning, and labor forecasting that rely on benchmark results will likely be misleading unless benchmarks are broadened or reweighted by economic task share. — If true, this changes expectations about near‑term automation, the shape of reskilling needs, and the validity of claims that AI will imminently displace large shares of labor.