A new study examines how large language models perform in a variety of medical contexts, including real emergency room cases — where at least one model seemed to be more accurate than human doctors.

The study was published this week in Science and comes from a research team led by physicians and computer scientists at Harvard Medical School and Beth Israel Deaconess Medical Center. The researchers said they conducted a variety of experiments to measure how OpenAI’s models compared to human physicians.

In one experiment, researchers focused on 76 patients who came into the Beth Israel emergency room, comparing the diagnoses offered by two internal medicine attending physicians to those generated by OpenAI’s o1 and 4o models. These diagnoses were assessed by two other attending physicians, who did not know which ones came from humans and which came from AI.

“At each diagnostic touchpoint, o1 either performed nominally better than or on par with the two attending physicians and 4o,” the study said, adding that the differences “were especially pronounced at the first diagnostic touchpoint (initial ER triage), where there is the least information available about the patient and the most urgency to make the correct decision.”

In Harvard Medical School’s press release about the study, the researchers emphasized that they did not “pre-process the data at all” — the AI models were presented with the same information that was available in the electronic medical records at the time of each diagnosis. 

With that information, the o1 model managed to offer “the exact or very close diagnosis” in 67% of triage cases, compared to one physician who had the exact or close diagnosis 55% of the time, and to the other who hit the mark 50% of the time.

“We tested the AI model against virtually every benchmark, and it eclipsed both prior models and our physician baselines,” said Arjun Manrai, who heads an AI lab at Harvard Medical School and is one of the study’s lead authors, in the press release.

To be clear, the study didn’t claim that AI is ready to make real life-or-death decisions in the emergency room. Instead, it said the findings show an “urgent need for prospective trials to evaluate these technologies in real-world patient care settings.”

The researchers also noted that they only studied how models performed when provided with text-based information, and that “existing studies suggest that current foundation models are more limited in reasoning over nontext inputs.”

Adam Rodman, a Beth Israel doctor who’s also one of the study’s lead authors, warned the Guardian that there’s “no formal framework right now for accountability” around AI diagnoses, and that patients still “want humans to guide them through life or death decisions [and] to guide them through challenging treatment decisions.”

In a post about the study, Kristen Panthagani, an emergency physician, said this is an “an interesting AI study that has led to some very overhyped headlines,” especially since it was comparing AI diagnoses to those from internal medicine physicians, not ER physicians.

“If we’re going to compare AI tools to physicians’ clinical ability, we should start by comparing to physicians who actually practice that specialty,” Panthagani said. “I would not be surprised if a LLM could beat a dermatologist at an neurosurgery board exam, [but] that’s not a particularly helpful thing to know.”

She also argued, “As an ER doctor seeing a patient for a first time, my primary goal is not to guess your ultimate diagnosis. My primary goal is to determine if you have a condition that could kill you.”

This post and headline have been updated to reflect the fact that the diagnoses in the study came from internal medicine attending physicians, and to include commentary from Kristen Panthagani.