Generative AI in French hospitals: medical secrecy, HDS and the CNIL

On a hospital ward, the temptation never goes away. A report to rephrase, a discharge letter to draft, a protocol to summarise for the night shift: a generative AI assistant does it in seconds. Many nurses, medical secretaries and managers have already tried it, often from a personal account.

The problem is what travels with the request. A name, a date of birth, a diagnosis, a room number. In France, this information is covered at once by medical secrecy, the GDPR and the rules on hosting health data. Here is what hospitals, clinics and health centres need to keep in mind.

Medical secrecy does not stop at the browser window

Article L.1110-4 of the French Public Health Code guarantees every patient respect for their private life and the confidentiality of information about them. This duty applies to every professional working in the health system, not only doctors. Breaching it is punishable under Article 226-13 of the Penal Code by one year in prison and a fine of 15,000 euros.

Pasting a patient file into a consumer chatbot means handing that information to a third party that is not part of the care team. The intention is good, saving time, but the act is still a disclosure.

HDS: where does the data pasted into a prompt actually live?

Article L.1111-8 of the Public Health Code requires anyone hosting health data on behalf of a healthcare facility or professional to hold HDS certification (hébergeur de données de santé, health data host). The certification framework was revised in 2024. It has applied to new certifications since November 2024, and hosts already certified had to comply before 16 May 2026.

The new framework requires, among other things, that data be hosted within the European Economic Area and that hosts tell their clients about possible access from third countries. When a staff member pastes a report into an AI tool, that data is processed and sometimes stored on the provider's servers. Nothing guarantees that the provider is HDS certified, or that the data stays in Europe. Facilities must therefore check each tool before any use with real data. We cover the location question in our article on sovereign AI for European CISOs.

What the CNIL and the HAS say

For the CNIL, France's data protection authority, health data are sensitive data under Article 9 of the GDPR: processing them is prohibited except in defined cases, and it requires stronger security measures. We set out the general obligations for companies in our first article on generative AI and the CNIL. In a hospital, the bar is higher: impact assessment, data minimisation, traceability of access.

On 30 October 2025, the Haute Autorité de santé (HAS), France's national health authority, published a guide called Premières clefs d'usage de l'IA générative en santé (first keys for using generative AI in health), aimed at professionals in health, social and medico-social care. It sums up its approach in four words: learn, verify, estimate, communicate. The guide stresses knowing the confidentiality rules and lists sharing confidential information among the poor practices it has observed.

The most common risky situations

Leaks almost never come from bad intent. They come from ordinary tasks repeated every day:

• rephrasing an operative report or a letter to the patient's GP;

• translating discharge instructions for a patient who does not speak French;

• summarising a file before a multidisciplinary team meeting;

• analysing a spreadsheet export containing names, social security numbers or lab results;

• drafting a reply to a complaint that recounts the care history.

In each case, the pasted text contains direct or indirect identifiers. Our article on the anatomy of an AI data leak shows how a single copy and paste is enough.

Ban it or govern it?

Blocking every AI tool is tempting, but rarely sustainable. Teams are short of time, and a ban pushes people towards personal phones, out of sight. A more realistic approach combines three elements:

1. A list of approved tools, chosen after checking hosting, contract terms and how the provider may reuse data.

2. A clear rule: no identifying patient data in an unapproved tool, even for a harmless task.

3. A technical control on the workstation that spots health data before it is sent and replaces it or blocks the request.

The third point makes the difference. A written rule does not protect an exhausted junior doctor at four in the morning. A tool that automatically swaps the name and social security number for fictitious values, while keeping the structure of the text, lets people keep working without exposing the patient. That is the idea behind format preserving anonymization.

Keep a record for inspections

In the event of a CNIL inspection, a certification audit or an incident, the facility must be able to show what it has put in place. A register of approved tools, a signed usage policy, incident logs showing what type of data was detected and what action was taken. This evidence matters as much as the measures themselves.

In short

Medical secrecy, the GDPR and HDS certification apply in full to generative AI assistants. Healthcare facilities do not have to choose between productivity and confidentiality: they need to choose their tools, set one simple rule and add a control at the moment data leaves the workstation.

ZeusLock detects and anonymizes health data in prompts and files before they reach ChatGPT, Copilot or Gemini, and logs every incident. Request a demo tailored to healthcare facilities.