Medical AI Surpassing Physicians

A Harvard study in Science found o1 achieved 67% diagnostic accuracy on emergency room triage cases compared to 55% for attending physicians—a 12-percentage-point advantage. A parallel study (Science.org, April 30) found o1-preview achieved 88.6% accuracy on clinicopathological cases versus 72.9% for GPT-4 and significantly lower for human clinicians. In oncology, a Northwestern study (Journal of Clinical Oncology, April 2026) showed Meta Llama 3.1 with DeepSeek generated more comprehensive pathology summaries than physicians. The most ambitious system, SPARK (Nature Medicine, May 2026), is an agentic framework autonomously generating cancer hypotheses across 5,400 patients and five tumor types. AlphaFold 3 underpins this pipeline with a 50% improvement in protein structure prediction and doubled accuracy for protein-ligand interactions.