Anonymize a médical report before ChatGPT: the guide

In many wards and practices, médical reports already go through ChatGPT or Claude to save time: summarizing a stay, rewriting a letter, structuring a file. The problem is that the report often leaves as-is, with the patient's name, their national ID, their dates and sometimes rare détails that make them recognizable. Yet health data is sensitive data under the GDPR (article 9), and médical confidentiality strictly governs its sharing. The good practice is not to ban AI, it is to anonymize the document before it leaves.

The patient and doctor identity is masked ; the clinical content (diagnosis, treatment) stays, so the AI actually helps.

Here, concretely, is what to mask in a médical report, how to do it while keeping the ability to restore the real values, and the mistakes to avoid.

What to mask in a médical report

A report holds far more identifying éléments than people think. The main ones:

• The patient's identity: name, first name, date of birth, family links mentioned.
• The national ID (social security number) and file identifiers.
• Sensitive dates: admission, birth, surgery, which let you cross-référence an identity.
• Addresses and contacts, for the patient and the correspondent.
• Rare éléments: an uncommon condition or a unique context can be enough to recognize someone.
• Practitioners and facilities, when they are not needed for the analysis.

The method, in three steps

• 1. Pseudonymize the report: each identifying élément is replaced by a neutral, consistent token (the patient becomes PERSONNE_1 everywhere).
• 2. Use the AI of your choice on the masked version: summary, rewriting, structuring. The AI never sees a real name.
• 3. Re-identify the result locally: tokens become the real values again, on your machine, with nothing sent back.

Pseudonymization or anonymization?

The two words do not mean the same thing, and the nuance matters. Pseudonymization is réversible: you replace values with tokens (PERSONNE_1, DATE_1), you keep a mapping key, and you can re-identify the AI's answer. Anonymization is irréversible: no way back. For a report you must re-read with the real names after the analysis, it is pseudonymization. Saying one for the other means getting your risk level wrong.

Mistakes to avoid

• Believing detection is perfect. No tool catches 100% of cases. A human review of the masked document is still needed before sending it.
• Forgetting attachments and scans. A scanned report or a screenshot contains text: you must OCR it then mask it, otherwise the leak goes through the image.
• Confusing blur with masking. Visually redacting is not enough if the text remains extractable behind it.
• Storing the key anywhere. The mapping must stay under your control, encrypted, separate from the document.

Safe-Doc pseudonymizes the report before the AI, keeps the layout (PDF and Word), and lets you re-identify the result locally. Processing happens in the European Union then is purged, and scans are rewritten via OCR. You keep the key. For sector-specific détails, see the Safe-Doc for healthcare page.