In Harvard study, AI offered more accurate diagnoses than emergency room doctors

The study's most consequential detail isn't the accuracy gap itself but which type of model cleared the bar. If a general-purpose LLM outperformed ER physicians, that directly undercuts the case for purpose-built clinical architectures and changes the build-vs-buy calculus for every health system currently evaluating specialized vendors.

Two days before this Harvard result published, The Decoder reported that Google DeepMind's specialized 'AI co-clinician' beats GPT-5.4 in blind physician tests but still trails experienced doctors. That framing now looks premature: if a general LLM surpasses ER physicians on real diagnostic cases, the argument for domain-specific architectures over general models weakens considerably, at least at the emergency triage tier. The ethical divergence benchmark covered the same day also matters here, because a diagnostic model that outperforms humans on accuracy but encodes inconsistent clinical ethics creates a liability surface that no hospital credentialing committee will ignore.

Watch whether the Harvard team releases a methodology appendix specifying which model won and whether the case set was prospective or retrospective. If the cases were drawn from historical records the model could have encountered during training, the accuracy advantage is suspect and the regulatory timeline argument collapses.