A lawyer opens Claude, protects a client folder with RAGbase Shield, and Claude reads the documents with codes instead of names. PERSONNE_A instead of the client. SOCIETE_A instead of the opposing company. It looks clean. But the prompt still carries a great deal. This piece looks at what stays readable once the mask is on, and why that matters under article 10 of the Paris bar’s model AI charter of July 2026.
Shield replaces names, companies, addresses and a list of identifiers with codes, on the lawyer’s own computer, before any text leaves for the model. This is local, reversible pseudonymisation. Shield is not anonymisation: it is a mask that can be lifted, on the same computer, once the answer comes back. That is the whole point, and also the whole limit. Shield masks the identity, not the secret itself.
What a masked prompt still contains
Once the mask is applied, the text that actually reaches Claude keeps the facts, the dates except dates of birth, the amounts, the legal reasoning, the courts, the laws and articles cited, the dossier code, the masked file names and the lawyer’s question. In other words, everything that makes the prompt useful to the model stays in the prompt. Only the names and numbers that point straight at a person or a company are swapped for codes.
This is where a project name becomes interesting. A code hides a name. It does not hide a fact. If the fact itself, a merger, an acquisition, a restructuring, a unique job title, a precise place, or a specific amount, points to one client and one client only, the mask does not change that. Facts can still identify a client even without a name. The CNB’s 2024 practical guide on generative AI systems (page 31) makes the same point: pseudonymise before the AI, replace names with initials or random strings, reduce addresses to the city, and take care with cross-references such as amounts and project names. Shield will not guess that a project name is a giveaway on its own. The firm has to spot such terms, and generalise them in the documents when needed.
Why the GDPR still applies
Article 4(5) of the GDPR, in the official text available at https://eur-lex.europa.eu/eli/reg/2016/679/oj, defines pseudonymisation as processing that stops data from being attributed to a person without additional information, kept separately and protected. Shield keeps that additional information, the code to name table, in a vault on the lawyer’s computer, encrypted, with the key stored in the operating system’s keychain. That design matches the definition. It does not change what the definition implies for the firm. For the lawyer, pseudonymised data remain personal data. The firm keeps the vault, so the firm keeps the ability to re-identify, and the GDPR keeps applying in full: a register of processing, a legal basis, information to the people concerned. Pseudonymising a prompt is not a way out of the GDPR. It is a way to meet it.
The Paris bar’s model AI charter of July 2026, in its articles 4 and 10, sets a similar bar for an online tool: it is allowed after validation or sufficient guarantees, without confidential or sensitive data, with systematic human review, and with security guarantees that include encryption and logging. Read against that text, a masked prompt that still carries a project name, a precise amount or a unique job title can still count as sensitive data, even though no name appears. The mask on names does not by itself satisfy the charter’s requirement to keep confidential or sensitive data out of the prompt. That requirement sits on top of the mask, and it is the lawyer who has to judge it case by case, not the tool.
What stays with the lawyer
Detection is never complete. Read the sensitive passages again, because a name on its own can be missed, and read the final document the same way a lawyer reads a letter before it goes out. Compliance with professional secrecy and with the GDPR remains the responsibility of the lawyer and the firm.
Shield is built first for French legal documents. Only a folder chosen in the Shield window is protected: a file attached, or text typed or pasted into the conversation, reaches Claude as it is. To try Shield, add it to the Claude app from the Install page, on Mac or on Windows.
The limits of Shield are on the Limits and compliance page.
Sources
https://cnb.avocat.fr/medias/cnb-guidepratique-utilisation-systemes-iag-2024-68f7b1d86e4570.98487114.pdf; https://eur-lex.europa.eu/eli/reg/2016/679/oj; https://ragbase.ai; docs/CONFORMITE.fr.md and README.md of RAGbase Shield.