On 6 January 2026, the US state of Utah crossed a symbolic line: it authorised an AI to renew routine prescriptions for chronic conditions.[2] A few weeks earlier, the French National Ethics Committee (CCNE) had organised a colloquium at ENS Paris entitled "What medicine in the AI era?"[4] Both events tell the same story: medical AI is no longer a distant promise. It is here. The real question is: how far will it go?
To be honest: in certain precise domains, AI already surpasses humans. In radiology, Google Health's algorithm analysed over 25,000 mammograms and detected breast cancer with 9.4% fewer false positives than experienced human radiologists.[1] In dermatology, neural networks match dermatologist performance on melanoma detection from photos. In cardiology, some algorithms predict heart attacks 24 hours in advance from ECGs that appear normal to the human eye.
In 2026, these tools are leaving the experimental phase and entering hospitals to assist caregivers and free up their time with patients.[6]
Artificial intelligence will never replace human diagnosis. Its goal has never been to substitute healthcare professionals, but to work alongside them.[5]
Why? Because medicine is not just about data. A patient coming to a consultation is not a set of symptoms to be optimised. They carry a history, anxieties, a family situation, values. Medical decision-making integrates all of this. AI, however powerful, cannot see the patient trembling, hear their hesitation, sense that they are hiding something.
In 2026, the question is no longer whether AI will transform medicine. It already is. The real question is more nuanced: what can and should a physician truly delegate to AI, and what remains irreducibly human in the act of care?
Here is the figure that should give pause: according to the DREES, 25–35% of physicians' time is devoted to administrative and documentary tasks.[3] Time stolen from patients. Time that fuels physician burnout. Time where nobody wins — neither the doctor nor the patient.
This is precisely where AI can — and should — intervene as a priority. Not to replace diagnosis, but to free physicians from tasks that do not require their human expertise: writing the consultation note, structuring the report, generating the referral letter, coding acts.
What medical AI concretely does in practice in 2026
Automatic consultation documentation · SOAP note generation · Drug interaction alerts · Referral letters · Prescribing assistance for chronic conditions · Detection of warning signals in lab results
In January 2026, Utah authorised an AI to renew routine prescriptions for stable chronic conditions — controlled diabetes, managed hypertension.[2] It is a world first. And it deeply divides the medical community.
The EU AI Act, entering full force in August 2026, settles the matter for Europe: medical AI systems are classified as "high risk" and require mandatory human supervision. No autonomous prescribing without medical validation. The line is clear.
No — if they know how to adapt. Yes — if they ignore what is coming. The medicine of tomorrow will not be medicine without doctors. It will be medicine where doctors do more of what drove them to choose this profession — caring for people — and less of what administratively exhausts them.
See also our articles on how to choose your medical AI and on liability in the event of AI-related medical error.
Can AI already make diagnoses?
Yes, in specific and delimited domains. In dermatology, algorithms match dermatologist performance on melanoma detection from photos. In radiology, Google Health's AI detected breast cancer with 9.4% fewer false positives than human radiologists on 25,000 mammograms. In cardiology, AI systems predict heart attacks 24 hours in advance from ECGs. But these performances concern very specific tasks on structured data — not the global medical diagnosis of a patient with their comorbidities, history and context.
Which doctors are most threatened by AI?
The most exposed specialties are those where work consists primarily of analysing visual or numerical data: radiology, anatomical pathology, dermatology (imaging). These specialties will not disappear, but their scope will transform — radiologists will increasingly supervise algorithms rather than directly reading each image. Conversely, specialties built on human relationships, touch and contextual decision-making (general practice, psychiatry, geriatrics) are far less exposed.
What do medical AI systems concretely do in practice in 2026?
The most deployed uses in 2026 are: automatic documentation (transcription and structuring of consultations), prescribing assistance (drug interaction alerts), detection of warning signals in biological data, and generation of referral letters. These uses free an average of 25–35% of physicians' administrative time according to the DREES — time returned to consultation and the patient relationship.
Can a physician be held liable for following an incorrect AI diagnosis?
Yes. Medical liability remains full regardless of the tool used. A physician who follows an AI diagnosis without exercising critical judgement engages their professional responsibility. The EU AI Act (August 2026) classifies medical AI as 'high risk' and requires mandatory human supervision. In practice, AI must be treated as a decision-support tool — like a lab result or imaging — not as a final decision.
Automatic documentation, SOAP notes, referral letters — so physicians spend more time with patients. Hosted in Europe, nFADP and GDPR compliant.
Try free →