ChatGPT does not heal. But if millions of French people confide their symptoms to him every day, it is perhaps less a question of technology than of a health system running out of steam.
Millions of French people now consult ChatGPT to describe their symptoms, before seeing a doctor, or instead. It’s understandable. 8 million people do not have a primary care doctor. In some under-resourced areas, getting an appointment with a GP takes six weeks. When the system is absent, the algorithm occupies the void. It’s rational. This is not, however, acceptable.
The market data is clear: in France, the majority of teleconsultations concern patients living in medical deserts. They reveal a simple reality: the demand for care is there, massive, and it is looking for a response. Generative AI has stepped into this gap. But it sometimes does so by producing a response that resembles medical advice, without assuming the requirements, responsibilities or guarantees.
Apparent competence, real risk
Language models are fluid, structured, convincing. They prioritize hypotheses, propose differential diagnoses with superficial precision. What they produce, technically, is a statistical prediction about the likely outcome of a medical text, not an assessment of the health status of a real person. They ignore family history, current treatments, drug interactions. Two patients with identical symptoms may require radically different care depending on their age or comorbidities. It is precisely this singular reading that the doctor operates, and that no algorithm can reproduce.
There is also a question of responsibility that tech cannot avoid. A doctor answers for each diagnosis before the Order, before the courts, before his patients. He is assured, he took an oath. An AI has neither an RPPS number nor professional liability insurance. In the event of an error, delayed heart attack, missed meningitis, emergencies overloaded by an unjustified alarm, no one responds. The consequences are real.
Performance on a typical case is not enough
Some will argue that comparative studies show diagnostic performance of AI equivalent to that of doctors on standardized cases. This work deserves to be taken seriously. They evaluate typical cases, under controlled conditions, without the relationship, without the context, without the capacity to perceive what the patient does not express. A lonely, worried patient who misinterprets an algorithmic response and postpones his consultation for three weeks: this is not part of any evaluation protocol. So the right question is not “is AI capable of diagnosing?” but “who benefits from the illusion that it can?”
AI at the service of doctors, and only on this condition
Recognizing that generative AI cannot replace the doctor does not amount to denying its potential. It’s even the opposite. The real issue is not whether AI will replace healthcare professionals, but how it can make them more effective for the benefit of patients.
Used in a medical setting, under the responsibility of healthcare professionals, AI can enrich clinical practice, reduce administrative tasks, improve care coordination and streamline patient journeys. Telemedicine is already demonstrating this. According to the Court of Auditors, it generates 113 million euros in annual savings for the health system. However, France still only carries out 4% of its medical procedures via teleconsultation, compared to 20 to 30% in the most advanced countries. AI integrated into these pathways, supervised and supervised by doctors, could amplify these benefits. On condition that it remains a decision-making tool, and not becomes a substitute for medical consultation.
This is where the real challenge lies. If millions of French people today ask ChatGPT about their symptoms, it is not because they prefer a artificial intelligence to a doctor. It’s firstly because they can’t find an available doctor. As long as this access remains insufficient, patients will continue to look elsewhere for answers.
The priority is therefore twofold: invest in access to care and set clear rules for the uses of AI in health. Patients must know how to distinguish information generated by an algorithm from real medical advice, and AI tools must be designed to assist professionals, never to delegate responsibility for diagnosis to them. Technology can help address the healthcare shortage. It must not become a replacement solution to this shortage.
