Diagnostic reasoning on clinical cases · Contested (since 2024)

In a randomized trial with 50 physicians on written cases, GPT-4 alone scores 92 percent on a diagnostic reasoning rubric, against 74 percent for physicians using conventional resources and 76 percent for physicians given the model.

These are vignettes: the history is already taken, the findings are already written down, and no one is examined. Later work (Microsoft's sequential-diagnosis benchmark on 304 NEJM cases, 2025) reports higher accuracy than physicians who were denied their usual tools, which is not how physicians work. Reading a case well is being contested in earnest; eliciting the history, examining the body, deciding under uncertainty with a real patient and answering for the outcome are not the same task, and remain with the clinician.