They cannot replace doctors: ChatGPT and similar tools make bad diagnoses

(Patria) - Artificial intelligence tools, such as the popular ChatGPT, are not good at making diagnoses, concluded a study whose results were published today, as the medical benefits of AI are closely examined worldwide.

The study, whose results were published in the journal Nature Medicine, conducted on 1,300 people in the United Kingdom, shows that several artificial intelligence models - ChatGPT, Llama (Meta) or Command R+ - do not perform better than a simple internet search when a patient is asked about their symptoms.

"There is great excitement about artificial intelligence models, but they are simply not ready to replace a doctor," said researcher Rebecca Payne from the University of Oxford and co-author of the study.

The study participants were not actually sick. They participated in a kind of role-playing game: researchers gave them ten different sets of symptoms for which there is a consensus in the medical community about the corresponding diagnosis.

Only a third of the participants received an accurate diagnosis. This is no better than the group that had to settle for a classic internet search.

However, some studies have shown that ChatGPT and other models can successfully pass medical exams: when placed in the role of a medical student, for example, facing multiple-choice questions, artificial intelligence systems perform well.

However, this new study shows that the situation changes when it comes to interaction with real people. They can be imprecise in describing their symptoms and not provide all the key elements, researchers say.

This study is part of broader discussions about the potential contribution of artificial intelligence to medical information, given that in many countries, including developed ones, there are often long waits to see a doctor.

In France, the High Authority for Health (HAS) is expected to soon issue an opinion on the benefits of direct use of AI for patients. At the end of 2025, it had already assessed that these tools could assist healthcare professionals if used reasonably.

The study published today has certain limitations, such as a methodology based on fictitious scenarios. The models used have also been largely replaced by more efficient versions in the meantime.

"But this is a very important study that highlights that chatbots pose real medical risks to the general public," bioethics expert David Shaw from Maastricht University in the Netherlands told AFP.