Reversible Anonymization: Use AI While Reducing Your Data's Exposure

Reversible anonymization replaces each sensitive value with a unique token before it reaches the AI. The mapping between the token and the real value stays on your side. After the model answers, you restore the real values. The result: the AI keeps its full meaning, but the real values don't leave. You keep the tool's usefulness. And you greatly reduce your data's exposure.

Reversible anonymization, in plain terms

The idea has three steps. One: we spot the sensitive data in your text. Two: we replace each value with a neutral token, for example [CLIENT_1]. Three: we keep a token↔value mapping. That table lives on your side, never with the AI provider. You send the text with tokens. The model replies with those same tokens. You swap them back for the real values on arrival. It's reversible: nothing is lost.

Why this beats deleting

The first instinct is often to erase. You cross out the name. You delete the number. But this method breaks the meaning. The AI no longer knows what you're talking about. A text full of holes gives a vague answer.

The reversible token avoids that trap. Let's compare the two approaches:

• Deleting breaks the thread: the AI loses context and answers off the mark.
• Deleting is final: the removed value can't come back into the answer.
• A token keeps consistency: the same client keeps the same token across the text.
• A token can be restored: you bring the real values back in the final answer.

The GDPR nuance, in full honesty

Let's be precise about the words. According to the CNIL, anonymization makes identification impossible, and in an irreversible way. Truly anonymous data then falls outside the scope of the GDPR. But a reversible token keeps a mapping. So you can go back to the real value. In strict GDPR terms, this is pseudonymization, not anonymization.

The difference matters. Pseudonymization (identifiers replaced by aliases, but reversible) remains personal data. It stays subject to the GDPR. So we will never say your data becomes « anonymous under the GDPR ». That would be false.

How ONYRI applies reversible anonymization

ONYRI Sanitize does this work in your browser. The engine spots the sensitive data, then replaces it with reversible tokens. The mapping stays in memory, client-side. Here's the flow:

1. 1The engine detects the sensitive data: built-in detectors plus your custom rules.
2. 2Each value is replaced with a token that's unique and consistent across the text.
3. 3You send the tokenized text to the AI, via the web app (Text, Tables, Chat) or the browser extension.
4. 4On the answer, ONYRI de-tokenizes client-side: the tokens become your real values again.

The mapping table never reaches any ONYRI server. A Free plan serves as an entry point to test the product. The Pro and Team plans add custom rules and volumes. To clarify the terms, read our comparison of anonymization, pseudonymization and tokenization.

Reversible anonymization offers a real balance. You keep the AI's usefulness. You keep your real values close at hand. Let's state the limit, in full honesty: this method is pseudonymization, not anonymization in the strict sense. It greatly reduces exposure, without making it zero. But it hands you control over what matters: your real data, which stays in your browser.