Healthcare AI's Oracle Problem: benchmarks can't prove that a model helps real patients
Scoring 100% on med-school tests doesn't make a doctor (or an AI) good at real patients, and that gap is the invisible bottleneck to healthcare AI. Evals in medicine are quietly subjective, trained on what physicians happened to do that day, and the data behind them is guarded and easily spun into marketing. The only honest measure requires tracking a patient forward over time, which is far harder than passing a benchmark. Buyers of hospital AI have no clean way to tell who's best.