How do you pseudonymise a document before sending it to ChatGPT, Claude or Gemini?

1 Start with the need, not the whole document

Before pseudonymising, define what you expect from the AI: summarise an argument, rephrase a clause, extract points, compare two versions, prepare an outline… The more precise the task, the easier it is to limit what you send. If three paragraphs are enough, there is no need to send fifty pages.

2 Spot direct and indirect identifiers

Names, addresses, emails, phone numbers or unique references are obvious identifiers. But context can also identify: a rare job title, a combination of date and place, a very specific amount, a project name or the description of a known event. Useful pseudonymisation is therefore not just a find-and-replace on names.

3 Replace with consistent aliases

Use stable aliases that preserve the logic of the document: for example PERS_1 and PERS_2 for two people. For other categories of information, apply a consistent naming scheme without reusing the same alias for two different entities. The same entity must keep the same alias throughout the text. This consistency lets the model understand relationships without needing to know real identities.

4 Keep the mapping separately

The table linking aliases to original values must not accompany the document sent to the AI. It must stay in a controlled and protected environment, because it allows re-identification. This separation is what gives pseudonymisation its meaning.

5 Do a final check before sending

Reread the pseudonymised version as if you did not know the file. Ask yourself three questions: is there any direct identifier left? Does a combination of information make someone easy to identify? Is any confidential information unnecessary for the task? If so, correct it before sending.

Depending on the context, you may need to go further: if the date, the job title or a circumstance makes the person easy to identify, these items must be examined too. Replacing names does not automatically turn the text into anonymous data.

6 Send the AI only the version it needs

The request must contain the pseudonymised text and the instructions useful for the task, not the mapping table. You should also use an account and a plan whose terms suit your professional use and obligations.

7 Restore only in your environment

If the result needs to become directly usable in the file again, the aliases can be replaced with the original information after the answer is received. This step must be done in the environment where the mapping is kept, without needlessly sending identities back to the AI provider.

The manual method and its limits

For a short text, this method can be done by hand. On long or repetitive documents, the main risks are missing an identifier, inconsistent replacements, wasted time and difficulty restoring the original values correctly. Human review remains necessary in every case.

How NYMETRA automates this flow

NYMETRA analyses the content in the client’s environment and highlights the information detected. The user can add, remove or correct items before sending. Approved items are pseudonymised, the necessary version is sent directly to the chosen AI provider, then the original information can be restored locally in the answer where the mapping allows.

Automated detection is not presented as exhaustive: the user approval step remains essential before every request.