A contract to get reviewed in French. A German report whose every clause you need to grasp, word for word. AI translates remarkably well, so the reflex is simple: you paste the whole file into the tool and wait for the result. Except that file holds the names of the parties, their addresses, amounts, sometimes a case number that is enough to cross-reference everything. All of it goes to the engine. In clear text. And yet, to translate a sentence, the engine never needed to know who is called what.
Anonymize
Replace names and data with consistent tokens.
Translate
The AI translates the text, the tokens stay unchanged.
Re-identify
Reinject the real values into the translation.
The trick: a pseudonym like [PERSON_1] is not a word to translate. It passes through translation identical to itself, so re-identification works just as well in the target language as in the source.
AI translates the content, not the identity
A translation engine works on meaning. On phrasing, too. It rewrites, shifts register, adapts an idiom that would make no sense taken word for word in the target language. The name "John Smith", however, does not get translated. It comes out "John Smith". Exactly as "Berlin" stays "Berlin" from one language to the next. Proper nouns pass through translation unchanged by nature. Replacing them with pseudonyms therefore takes nothing away from the engine's job: it translates the same sentence, with a neutral subject in place of the real name.
Why a pseudonym does not bother the translator
A consistent pseudonym like PERSON_1 is not a vocabulary word. The engine copies it as-is. The structure of the sentence, the meaning, the tone: all of it stays intact, because nothing pushes the AI to touch a string it does not recognise as a translatable word. And because the pseudonym comes out identical on the translated side, the final re-identification works just as well in the target language as in the source. You get back a translated text, with the real values. Without any nominative data ever leaving in clear text.
The frequent case of cross-border documents
This is the most common situation. A French-German contract. Terms and conditions to be issued across several countries, an audit report bound for a foreign parent company: so many documents that move between jurisdictions and, by design, concentrate personal data. Handing them to a third-party AI without precaution simply means exporting that data. By pseudonymizing first, you get the same translation, usable as-is. And you keep control of the identities.
Safe-Doc masks the document before AI translation, then re-identifies. The layout, meanwhile, stays put. To go further, see the ChatGPT and GDPR at work guide.
Translate without exposing. Pseudonyms on the way out. Real values on the way back.