Harvard Medical School Study Finds AI Outperforms Doctors in Emergency-Room Diagnosis

Emergency-room triage requires clinicians to identify possible causes with limited information and under severe time pressure. Misdiagnoses can delay treatment and drive up healthcare costs. Harvard Medical School and Beth Israel Deaconess Medical Center therefore compared OpenAI’s o1 and GPT-4o with attending physicians to assess the value of LLMs as clinical decision-support tools. The study did not estimate potential cost savings and stressed that AI cannot replace doctors.

Published in Science on April 30, 2026, the study used text-based medical records from 76 real emergency-room patients in a blinded evaluation. Across the three stages—initial triage, physician assessment, and admission to a hospital ward or intensive care unit—o1 produced correct or closely matching diagnoses in 67.1%, 72.4%, and 81.6% of cases, respectively. The corresponding rates for the two attending physicians were 55.3% and 50.0%, 61.8% and 52.6%, and 78.9% and 69.7%.