OpenAI model outperforms doctors at diagnosis, but researchers urge caution

For the first time, an OpenAI language model has outperformed doctors in diagnostic accuracy on real emergency-room data. The study was published in the journal Science on April 30.

What the study showed

OpenAI's o1-preview model analyzed medical records from 76 real cases in an emergency department. At different stages of care (on admission, after the physician's examination, after transfer to another department), the model produced a diagnosis in parallel with two physicians. And it guessed right more often: at the final stage, 82% of its diagnoses were correct, versus 79% and 70% for the doctors.

Interestingly, both the humans and the model did better when more information was available. But the AI kept its lead at every stage, even with incomplete data.

• 82% diagnostic accuracy versus 79% and 70% for the doctors
• Tested on real emergency-room records
• The model analyzed full sets of details
• Its results improved with each new piece of information

But the medical community is cautious

The authors themselves are quick to clarify: AI does not replace doctors. "I don't think our results mean that AI will push out doctors," says co-author Arjun Manrai of Harvard Medical School. His colleague Adam Rodman, a medicine instructor in Boston, adds: "The results are great, don't get me wrong, but I'm slightly worried about how they might be used."

The main problem is that there is no single standard for how to evaluate LLMs on medical tasks at all. Some researchers count it as a success if the model named 5 of 7 possible diagnoses. Others see the same result as a complete failure. The same outcome is judged differently.

The reliability problem with chatbots

Parallel studies show that chatbots often make things up when answering medical questions. Almost half of the answers contain errors: invented sources, inaccurate advice, falsehoods delivered with confidence. The model sounds equally convincing whether it is right or not.

"These models are used every day, and there is a certain risk that no one is measuring or reducing," —

Arya Rao, Harvard

For a doctor the task is harder: when a model offers a consultation, the physician has to quickly tell whether it is correct or a hallucination. A doctor will, of course, be better placed to judge which information matters. But spotting a lie inside a convincing answer is a challenge.

What this means

OpenAI has already launched ChatGPT for doctors and for healthcare. The technology is moving faster than medicine can regulate and test it. What is needed are real clinical trials and clear workflows in which the doctor uses AI as an aide in consultations, not as the final answer.

The speed of innovation matters, but accountability is needed more.