Medical confidentiality is not optional. The law protects it, and it covers everything a professional learns from a patient, from the first consultation right through to the last letter filed in the record. A report dropped into ChatGPT to save five minutes? That is the confidentiality now at stake. The answer is not to ban AI. It is simple: never show it what identifies the patient.
Patient document
Report, letter, file.
Safe-Doc
Identifying data becomes pseudonyms.
AI
ChatGPT, Claude, Gemini work on the masked text.
Re-identify
Real values return locally, key kept.
What médical confidentiality protects
So what does it actually cover? The patient's identity, their health status, the care provided, and more broadly everything that is confided or observed in the consulting room, down to the smallest note. Sharing it with a third-party service, even for a quick summary, already steps outside the care framework. Pseudonymizing before the AI changes that. You keep the tool. You do not expose the patient.
The method, without breaking confidentiality
The diagram above says it all. You pseudonymize the document, you work with the AI on the masked version, then you re-identify the result locally, on your own machine, without ever letting the mapping leak. At no point does the AI see a real name. Or a real date. Or a real national ID.
The limits to know
No detection is perfect. Always review the masked document before sending it, because a thirty-second read beats a stray patient detail slipping out to a third-party model you do not control. Scans hide text: OCR it, then mask it. The mapping key, for its part, stays encrypted and under your control alone.
Safe-Doc pseudonymizes before the AI, preserves the layout, processes your documents in the European Union, then purges everything once the work is done. The rest lives on the Safe-Doc for healthcare page.
Try it on a real report. Reversible pseudonymization. Confidentiality preserved, end to end.