ChatGPT and medical secrecy: what a doctor can share with AI

What does French law say about medical secrecy and online AI?

No French statute mentions ChatGPT, but four sets of rules apply as soon as a doctor pastes a report into an online AI tool.

Secrecy covers everything the doctor saw, heard or understood

Article L1110-4 of the French Public Health Code (Code de la santé publique) sets out the principle: anyone cared for by a health professional or facility has the right to respect for their private life and to the secrecy of information about them. That secrecy covers all information about the person that comes to the professional's knowledge, and it binds every professional working in the health system.

The French code of medical ethics, codified as Article R4127-4 of the same code, adds that secrecy covers everything the doctor learned in the course of practice: not only what was confided, but also what the doctor vu, entendu ou compris (saw, heard or understood). A diagnosis, a working hypothesis, a remark about the patient's family: all of it is covered, not just the name.

Sharing is tightly framed: Article L1110-4 only allows exchanges between professionals who all take part in the patient's care, and only for what is strictly necessary for coordination or continuity of care. An AI provider is not one of them.

Finally, Article 226-13 of the French Criminal Code punishes the disclosure of secret information by the person entrusted with it with one year in prison and a €15,000 fine. A data processing agreement (DPA) with the provider deals with the GDPR, not with secrecy.

Health data, a special category under the GDPR

The GDPR defines data concerning health as personal data related to a person's physical or mental health, including the provision of health care services, which reveal information about their health status (Article 4(15)). A consultation report is full of it.

Article 9 makes it a special category: processing is prohibited, with listed exceptions. One covers medical diagnosis and the provision of care, where the data is processed by or under the responsibility of a professional bound by secrecy (Article 9(2)(h) and 9(3)). Sending a file to an AI service is one more data flow that the practice or hospital, as controller, must be able to justify. More detail: ChatGPT and GDPR: can you put client data into it?

HDS-certified hosting: is a chatbot a host?

Article L1111-8 of the Public Health Code applies to anyone who hosts personal health data collected during prevention, diagnosis, care or social and medico-social follow-up, on behalf of those who produced or collected it, or on behalf of the patient. For digital data, the host must hold a certificate of conformity issued by an accredited certification body: this is HDS certification (hébergeur de données de santé), whose framework is published by France's digital health agency, the Agence du numérique en santé.

An AI service that keeps conversation history does, in fact, store whatever was pasted into it. Does that make its provider a “host” under Article L1111-8? That is a question of legal characterization, depending on the service and the contract, to settle with your DPO. The precaution, on the other hand, is not in doubt: if the copy you send no longer contains the patient's identity, most of the question falls away.

What the HAS, the CNIL and the medical council say

The HAS. Its guide Premières clefs d'usage de l'IA générative en santé (available in English as “Initial key guiding principles for the use of generative AI in healthcare”), adopted on October 23, 2025 and updated on April 15, 2026, is built around four principles, A.V.E.C. (Apprendre, Vérifier, Estimer, Communiquer). Its key 2.4, ne pas partager d'information confidentielle (do not share confidential information), says that whenever a generative AI system is available online, or without assurance that confidentiality is respected, professionals should check in every prompt that no information allowing direct or indirect identification, or covered by medical secrecy, is shared. It lists what to remove: last names, first names, initials, dates, addresses, phone numbers and social security numbers.

The CNIL. In its Q&A on using a generative AI system (July 18, 2024), France's data protection authority advises never sharing confidential information such as personal data with a consumer service, and favoring on-premise solutions for client or staff data. With the HAS, it put a draft guide on supporting the appropriate use of AI systems in care settings out for public consultation until April 16, 2026; the final version had not been published as of this article.

The medical council. As of October 21, 2026, we found no recommendation from the Conseil national de l'Ordre des médecins, France's national medical council, specific to generative AI. The code of medical ethics applies as it stands, starting with Article R4127-4 above.

What can you share with ChatGPT, and what must you keep?

The real question is not “ChatGPT or not” but “with what data”. Three families of uses, from the simplest to the one to rule out:

A medical decision remains the doctor's. The HAS says every use of generative AI should be conscious, supervised and reasoned, and OpenAI itself states that ChatGPT Health “is not intended for diagnosis or treatment.”

Which AI offerings exist for health professionals?

What the providers publish, as checked on October 21, 2026. These terms change often; check the pages listed in the sources.

• ChatGPT individual plans (Free, Go, Plus, Pro). Conversations may be used for training unless “Improve the model for everyone” is turned off, and there is no DPA. Business and Enterprise: no training by default and a DPA, which settles part of the GDPR question, not medical secrecy.
• ChatGPT Health. A space where users discuss their own health and can connect apps or, in the US, their medical records. According to OpenAI's Help Center, it is available to eligible, logged-in Free, Go, Plus and Pro users “in the U.S.”, and early access excluded the European Economic Area, Switzerland and the UK. It is not a tool for processing your patients' files.
• ChatGPT for Healthcare. An enterprise version for clinicians, administrators and researchers, with a Business Associate Agreement (BAA) for HIPAA-compliant use and no training on content. HIPAA is a US law; the page mentions neither France, the GDPR nor HDS certification.
• French medical assistants. MedGPT, for instance, is developed by Synapse Medicine for health professionals. Its privacy policy (March 2026) lists, among its hosting providers, “Microsoft Azure, hébergeur HDS” (an HDS-certified host): this is the publisher's claim, which we have not verified. Above all, the same policy and the service's home page state that no identifying patient data should appear in questions (aucune donnée identifiante relative aux patients).

So even tools built for clinicians ask you to keep the patient's identity out. The full comparison (ChatGPT, Claude, Mistral, Copilot): AI offerings, plan by plan.

How do you pseudonymize a medical report before using AI?

Pseudonymizing a report means replacing the identity of the people in it with stable labels, such as [PERSONNE_1], while keeping the medical content readable. The method with AnonymFiles, pseudonymization software for healthcare that does everything on your computer, with no account and no telemetry:

1. Drop in the file. Reports and letters in Word or PDF, scanned results and prescriptions (read through local OCR, see anonymizing a scanned PDF), .eml or .msg emails, .xlsx, .xls or .csv tracking sheets. The same patient keeps the same label across every document.
2. Let detection do its work. Masked automatically:
   • the names of the patient, relatives and colleagues;
   • addresses, emails and phone numbers;
   • the French social security number (NIR), recognized by its structure;
   • the date that follows né le, date de naissance or DDN (born on, date of birth);
   • numbers introduced as such: RPPS and ADELI practitioner numbers, FINESS facility numbers, IPP or patient file number, INS (national health identity), stay numbers (NDA, IEP) and health insurance membership numbers;
   • in a spreadsheet, columns headed Patient or social security number, in full.
   Medical vocabulary stays readable: in French, an IPP de 15 % (a 15% permanent partial disability rating) has no number and is never masked.
3. Review side by side. Original on the left, copy on the right. One click masks what detection missed, or what points to the patient without naming them: a rare condition, an admission date, a highly specialized department. The re-identification risk indicator (low, medium or high) flags what remains in the clear, including common health mentions (hospitalization, long-term condition, disability…), full dates and small towns. Nothing is exported until you approve it.
4. Export for the AI. The app produces the labeled copies, with metadata removed, and an instruction to paste so the AI keeps the labels intact.
5. Put the real names back. Paste the AI's answer: the app restores the real names on your computer, as text, Word or PDF. The encrypted mapping table (AES-256-GCM) never leaves the computer: see the mapping table.

Three sample prompts to paste after the app's own instruction; always review the answer.

Chronological summary of a file

Here are the pseudonymized documents from one patient's file. Labels in square brackets, such as [PERSONNE_1] or [DATE_1], replace masked data: keep them exactly as they are. Draw up a chronological summary as a table: date, source document, clinical event, current treatment. Add nothing that is not in the documents and flag any inconsistencies between them.

Discharge letter

Using this pseudonymized hospital report, draft a discharge letter to the patient's GP, [PERSONNE_2]: reason for admission, main findings, discharge medication, follow-up required, scheduled appointments. Plain, professional style, one page at most. Keep the bracketed labels exactly as they are and invent no data.

Explanation for the patient

Explain this report to [PERSONNE_1] in plain language, without jargon, in five to eight sentences: what was done, what was found, what happens next. Add no diagnosis or advice that is not in the document. Keep the bracketed labels exactly as they are.

On our corpus of fictitious French health documents, detection finds 34 items out of 34, with none of the 6 terms that must stay readable masked by mistake. It is our own corpus, built by us: it is no substitute for a test on your own files. Method and limits: our published measurements.

Is pseudonymization enough? The re-identification risk

Removing the name is not always enough. In healthcare, indirect identifiers are everywhere:

• a rare condition, especially combined with a region;
• age, town and occupation, when the town is small or the occupation rare;
• exact dates of admission, accident or surgery;
• a highly specialized department, or the name of a colleague who is the only one in their specialty locally.

This is the “indirect” identification the HAS refers to. Review is where you mask or generalize what points to the patient (“a town in the Ain department” rather than the village name, “in spring 2025” rather than the date). More detail: the re-identification risk.

Finally, a pseudonymized copy is not anonymous: as long as a mapping table can lead back to the names, it remains personal data under the GDPR (Recital 26). But the identity stays with you, and the table never leaves your computer. Further reading: anonymization vs. pseudonymization under the GDPR.

What about research, registries and medical expert reports?

Research and registries. Health research in France has its own framework. The CNIL has adopted reference methodologies: for research with consent (MR-001), a project that strictly complies only requires a declaration of conformity with the CNIL; otherwise, an authorization request is filed after an opinion from the relevant ethics committee. Pseudonymizing extracts before analysis helps minimize data; AnonymFiles prepares the documents, but it replaces neither the formality, nor the impact assessment, nor patient information.

Expert reports and personal injury. Medical experts and insurers' medical advisers handle thick files, often scanned; AI helps build a timeline or outline a report, on a pseudonymized copy. The French abbreviation IPP shows the difficulty: IPP : 123456 is a patient identifier, masked; IPP de 15 % is a disability rating, left readable. Lawyers who receive these reports have their own page: AnonymFiles for lawyers.

In short

ChatGPT and medical secrecy are not incompatible, on one condition: the patient's identity never leaves the practice. Pseudonymizing a report on your computer, reviewing the copy, then checking the AI's answer is the simplest method to apply and to justify. AnonymFiles costs €290 excl. VAT per seat per year, €240 from 3 seats, with a 14-day trial: see the healthcare page and pricing.