Physicians, AI and health data: GDPR Art. 9, HDS, medical secrecy
Last updated : 2026-10-11
Health data is sensitive data (GDPR Art. 9); the French requirement for a certified health data host (HDS, Article L.1111-8 of the Public Health Code) targets whoever hosts that data on behalf of the professional or the patient, which does not describe processing that stays on the physician’s own computer. That answer does not settle everything: the text sent afterwards to the model provider remains subject to the GDPR and to medical secrecy.
What the GDPR says
Article 9(1) of the GDPR (Regulation (EU) 2016/679 of 27 April 2016) prohibits the processing of certain categories of data, including “data concerning health”. Article 4(15) defines them as personal data related to the physical or mental health of a natural person, including the provision of health care services, which reveal information about their health status.
The prohibition has exceptions. Article 9(2)(h) covers processing necessary for medical diagnosis, the provision of health care or the management of health care systems. Article 9(3) adds a condition: the data must be processed by or under the responsibility of a health professional subject to professional secrecy, or by another person also subject to an obligation of secrecy.
What medical secrecy says
Article R.4127-4 of the French Public Health Code (the medical code of ethics, in force since 8 August 2004 according to Légifrance) states that professional secrecy, established in the interest of patients, binds every physician under the conditions set by law. The commentary of the French National Medical Council explains that the secret covers everything the physician learned in the course of practice: what was confided, but also what the physician saw, heard or understood.
What HDS is, and when it applies
HDS stands for “hébergeur de données de santé” (health data host). Article L.1111-8 of the Public Health Code (version in force since 1 July 2025 according to Légifrance) covers any person who hosts personal health data collected during prevention, diagnosis, care, or social or medico-social follow-up. That hosting is done on behalf of the persons who produced or collected the data, or on behalf of the patient. For digital data, the host must hold a certificate of conformity issued by an accredited body.
The HDS certification standard v2.0 of the Agence du Numérique en Santé (April 2024, order of 26 April 2024) sets the scope. Certification applies to anyone who provides all or part of a hosting service and acts as a processor within the meaning of GDPR Article 28. Six activities are covered: physical sites, hardware infrastructure, virtual infrastructure, the application hosting platform, administration and operation of the information system, and backup of health data.
These texts describe a provider that hosts on behalf of a third party. They do not mention the workstation of a physician who processes their own data. The draft HAS-CNIL guide (working document of 16 February 2026) recalls that when health data is hosted on a third party’s servers, for example in the cloud, that third party must be a certified health data host. Whether sending a text to an AI service falls under that rule depends on the service, its contract and the data sent. The texts cited here do not answer that question for a consumer AI service: ask your DPO or a lawyer, and ask the provider.
This certification is specific to France. In other EU countries, GDPR Article 9 applies, but hosting and secrecy rules come from national law.
What ndaact.ai does, and its limits
ndaact.ai detects personal data in the text on your computer before it is sent (names, dates of birth, identifiers, health mentions) and replaces it with markers. The model provider receives the markers in place of the detected entities. The mapping table that lets you read the answer stays on your machine. This is pseudonymisation within the meaning of GDPR Article 4(5): for you, the data remains personal data.
Detection has limits you need to know:
- Health data has no fixed form. A diagnosis, a treatment or a history can be written in countless ways. The model’s detection is probabilistic and can miss mentions, which are then sent in clear text.
- Indirect identifiers remain a blind spot. The draft HAS-CNIL guide notes that a patient’s disease history can become an identifying piece of data, notably for rare diseases. A freely written clinical case can allow the patient to be found by cross-referencing, even without a name.
- The redacted text still goes to the model provider. Markers remove neither the clinical context nor anything that was not detected.
- Read the text before sending it. The tool is an aid, not a guarantee.
What this guide does not say
ndaact.ai is one technical measure among others. Using it does not make processing GDPR-compliant, does not remove the need for an impact assessment where one is required, does not replace a contract with the provider and is not an opinion from your professional body. The decision to send text taken from a patient file to an AI service remains a professional decision that is yours to make.
This page is a translation. The French version prevails.