Harvard Medical School tested what happens when an AI diagnoses real Emergency Room patients.
A patient walked into Beth Israel Deaconess Medical Center in Boston with routine respiratory symptoms. Recent organ transplant. Immunosuppressed. The doctors treated it as a standard case.
The AI flagged a flesh-eating infection.
12 to 24 hours before the human physicians became suspicious, the AI had already identified it. The patient needed emergency surgery. The AI saw it first. From nothing but the text in the chart.
This is not hypothetical. It is published in Science, April 30, 2026. The lead authors are Peter Brodeur and Thomas Buckley. The co-senior authors are Arjun Manrai at Harvard Medical School and Adam Rodman at Beth Israel. The model they tested was OpenAI o1.
They ran it on 76 real ER patients at Beth Israel. Not textbook cases. Real people who walked into a real hospital.
At triage, when information is sparsest, the AI identified the correct diagnosis 67.1% of the time. The first attending got 55.3%. The second attending got 50.0%. It used only the text in the electronic health records. No imaging. No physical exam. No bedside manner. Just the notes.
On management decisions, the AI scored a median 89%. Physicians using UpToDate and their full clinical references scored 34%.
On the 100 most challenging diagnostic cases ever published by the New England Journal of Medicine, the AI included the correct diagnosis nearly 80% of the time.
As of 2025, 1 in 5 doctors and nurses worldwide already use AI for a second opinion on complex cases. More than half want to.
Here is where the fight starts.
A Harvard Medical School researcher not involved in the study said this. "When we say clinical reasoning, it doesn't mean the same thing as model reasoning. These models have been optimized to do sequential thought that we call reasoning, but it's not at all the same thing as how we teach medical students to reason."
The AI does not feel the pulse. It does not hear the hesitation in a voice. It does not notice the thing the patient did not say. It processes text.
And it was more accurate than the humans who wrote that text.
Arjun Manrai, the lead Harvard researcher, said it plainly. "We're witnessing a really profound change in technology that will reshape medicine." Then he added the line the AI cannot say for itself. "Humans want humans to guide them through challenging treatment decisions."
The question is no longer whether AI can diagnose you. It can. The question is whether you want it to.