The EDPB guidelines 01/2025 on pseudonymisation, adopted on 16 January 2025 and submitted for public consultation, set out a principle that sounds simple and is not: pseudonymised data remain personal data, and the thing that allows someone to go back to the real identity, the key or the mapping table, must be kept apart from the pseudonymised data itself. The guidelines link this separation to the general duties of the GDPR, articles 5, 25 and 32, on data minimisation, data protection by design, and security of processing. This piece describes the text submitted for consultation.
For a lawyer working with Claude, this separation principle is not an abstract idea. It describes exactly the boundary that decides whether a client’s name has left the office or not.
What separation means for a legal AI workflow
A pseudonymisation system has two parts. One part is the text that carries the codes: PERSONNE_A, SOCIETE_A, RG_1. The other part is the table that says PERSONNE_A means Mr so-and-so. The EDPB guidelines treat the second part as the secret. If that secret sits next to the pseudonymised text, or travels with it, separation has failed, whatever the label on the box.
RAGbase Shield is built around this separation. Each dossier has its own encrypted vault, a file combining SQLite storage and Fernet encryption, and the encryption key lives in the Mac keychain, or in the Windows Credential Manager. The vault and the key stay on the lawyer’s own computer. They are never part of the text that goes to Claude. What Anthropic, the provider of Claude, receives is the masked text: codes, dates except dates of birth, amounts, legal reasoning, the dossier code, nothing that opens the vault.
This matches the EDPB’s language, where the guidelines describe pseudonymisation as conditional on the secret being protected so that re-identification stays under the control of the party who holds it. Shield masks the identity, not the secret itself.
Shield’s local engine works from rules, dictionaries and check digits, in French and in English. It runs on the lawyer’s computer and makes no network call.
Codes without a name inside them
One detail in the EDPB guidelines matters more than it looks: the pseudonym itself must not leak information about the identity it replaces. A code derived from the name, or predictable from it, weakens the separation even if the table is well protected.
Shield allocates codes in order, PERSONNE_A, PERSONNE_B, SOCIETE_A, never derived from the name they stand for. Inside one dossier, every mention of the same person gets the same code, whether the document calls them by full name, surname first, or title and surname. The code carries no trace of the name itself, so reading the masked text alone tells the AI nothing about who PERSONNE_A is. The link between the code and the name sits only in the vault, and the vault sits only on the local machine.
What this does and does not solve
Separation protects the mapping table. It does not turn the pseudonymised text into anonymous text. For the lawyer, pseudonymised data remain personal data. A masked file still carries facts, dates, amounts, legal reasoning, courts, laws and the dossier code. None of that is anonymous under the GDPR’s own definitions, and the EDPB guidelines are explicit that pseudonymisation is a risk-reduction measure, not an exemption from the rules that protect personal data.
Facts can still identify a client even without a name. A well-known case, a unique job title, a precise place, an amount, all of these can point back to a person even once the name itself is gone. Shield’s architecture addresses the name and the identifiers it is built to recognise. It does not evaluate whether the remaining facts, taken together, are unique enough to identify someone. That judgment stays with the lawyer, case by case.
There is also a second secret that separation does not touch: professional secrecy itself, under article 66-5 of the law of 31 December 1971 and article 2 of the RIN. Shield masks the identity, not the secret itself. A fact covered by professional secrecy stays covered by it after masking. Replacing a name with a code does not lift the lawyer’s duty over the underlying fact, whether that fact is a health condition, a financial arrangement or a pending dispute.
And separation, however carefully built, cannot undo a mistake made before the tool acted. Text already sent cannot be recalled. Raw text can reach the AI provider before Shield acts: a file attached to the conversation, or a folder added to it instead of chosen in the Shield window. No vault and no key separation after the fact changes what the provider already received.
Where this leaves the firm
This piece is not legal advice. Compliance with professional secrecy and with the GDPR remains the responsibility of the lawyer and the firm. A firm’s own risk analysis decides whether a separation design meets its obligations under articles 5, 25 and 32.
The vault that holds the mapping between codes and real names is kept apart, on the local machine, encrypted, from the text that leaves for an AI provider. That separation is the EDPB’s own test for whether pseudonymisation does the work the regulation expects of it.
The limits of Shield are on the Limits and compliance page.
Sources
EDPB, Guidelines 01/2025 on pseudonymisation; docs/CONFORMITE.fr.md and README.md of RAGbase Shield