Everyone talks about pseudonymization before AI. But how does it actually work? The mechanism unfolds in three steps: replace identifying data with tokens, keep on your side the key that links each token back to its real value, then restore those values once the AI has finished. Three steps. Let's walk through them.
1. Replace with tokens
A name. An email. An IBAN, a phone number. Each piece of identifying data is replaced by a neutral, consistent token, so that Jean Dupont becomes PERSON_1 everywhere in the document, without a single exception. The meaning stays. So does the structure. Only the identity disappears.
2. Keep the key
The mapping table is the key. It stays with you, locally, and it alone knows that PERSON_1 stands for Jean Dupont. It is never sent to the AI. Without it, the tokens say nothing: noise, and nothing more.
3. Re-identify
The AI replies. Summary, rewrite, analysis: it makes no difference. You then apply the key in reverse to reinject the real values exactly where the tokens were, and you get back a result you can use right away. The AI, for its part, never saw a single real data point.
Safe-Doc runs these three steps for you, automatically. The key stays on your side. Processing happens in the European Union, then everything is purged. To go further, see the ChatGPT and GDPR at work guide and the detailed difference between pseudonymization and anonymization.
Run the test on a real document. Tokens on the way out, real values on the way back.