Words matter. A tool that promises anonymity promises that nobody can ever find the person again. RAGbase Shield makes no such promise. It does local, reversible pseudonymisation. On your own computer, it replaces client names and numbers with codes. Only then does the text go to the AI. The real names come back only on your computer, in the final document that Shield saves in the “Shield - sorties” folder.
This post explains the difference and what it changes for a law firm. It is based on the documentation of Shield: the README, the privacy model and the French compliance sheet.
What Shield does, on an example
Here is a fictional example. It is in French, because Shield is built for French legal documents first:
M. Thibault Mercadier, domicilié 12 rue des Glycines, 69007 Lyon, conteste son
licenciement par la société Ateliers Brunelac SAS, notifié le 3 mars 2025.
The AI receives this text:
M. PERSONNE_A, domicilié ADRESSE_1, conteste son
licenciement par la société SOCIETE_A SAS, notifié le 3 mars 2025.
The name, the address and the company became codes. The date stayed readable. Inside one dossier, a person always keeps the same code. The table that links each code to a real name is called the “vault”. There is one vault per dossier. It is encrypted and stays on your computer. Its key sits in the system keychain.
Why it is pseudonymisation
The GDPR defines pseudonymisation in article 4(5). The data can no longer be attributed to a person without additional information. That information must be kept separately and protected.
With Shield, the additional information is the vault and its key. They stay on your computer, apart from the text that is sent. Codes are allocated in order (PERSONNE_A, then PERSONNE_B). They are never computed from the name. So a code says nothing about the person to anyone who does not hold the vault.
Why not anonymisation?
There are two reasons.
The link exists, on purpose. The vault links each code to a real name. That link is what lets Shield put the real names back into the draft submissions. True anonymisation breaks that link for good. Shield keeps it, on your computer, for you.
Facts stay readable. It is not anonymisation: facts, dates and amounts stay readable by the AI. The AI needs them to do the work. It must see the dismissal date, the amount claimed and the court. Shield does not mask courts, institutions, laws, codes, articles, months and days either.
What it changes for the firm
The GDPR still applies
For you, the holder of the vault, pseudonymised data are still personal data. Recital 26 of the GDPR says so, and the EDPB recalls it in its guidelines 01/2025 on pseudonymisation. That text was adopted on 16 January 2025, in a version submitted for consultation.
Your obligations stay the same: record of processing, legal basis, information of the people concerned. Shield is a protective measure. It does not lift those obligations.
Facts can be enough to recognise a client
A masked name is not always enough. A well-known case, a unique job title, a precise place or an amount can identify a client. Shield does not mask facts. Read again, and generalise when needed.
Professional secrecy does not go away
Shield masks identity, not secrecy. A fact covered by professional secrecy stays confidential after masking. The compliance sheet says so about the Paris bar’s model AI charter (July 2026, articles 4 and 10). Validating the AI tool and the human review remain the firm’s job.
Recommendation no. 2026-01 of 2 April 2026, from the ethics board of the lawyers at the Conseil d’État and the Cour de cassation, points the same way. It lists pseudonymisation among the means to use. It also recalls that a legal question asked in abstract terms does not breach secrecy.
The risk goes down, it does not go away
The Shield compliance sheet keeps two points from the CNB reading grid on AI (v1.2, June 2025): a vendor’s promise not to reuse data cannot be checked, and only complete prior anonymisation or a local AI removes the risk.
So the Shield position is simple. Shield reduces what the AI provider receives. It does not remove the risk, because the facts still leave. To remove it, you need an AI that runs on your own machines.
Three rules to give the team
- Choose the folder in the Shield window. Do not add it to the conversation. A file attached or a folder added to the conversation reaches Anthropic before Shield can act.
- Do not paste raw text into the chat. Text typed or pasted straight into the conversation is not masked. To talk about a person, write their code (PERSONNE_A).
- Text already sent cannot be recalled. Shield protects only what goes through it.
What the firm can show
The firm can show a compliance lead or a DPO:
- the open source code, under the Apache-2.0 licence, public on GitHub;
- the fact that everything runs on the computer: Shield sends nothing to RAGbase;
- the encrypted vault, one per dossier, which stays on the computer with its key;
- the local journal, with events, codes and counts, never a real name.
In short
| Question | Answer |
|---|---|
| What does Shield do? | Local, reversible pseudonymisation: codes instead of names, an encrypted vault on your computer. |
| Are the data still personal data? | Yes, for the firm that holds the vault. The GDPR applies. |
| Are facts masked? | No. The AI needs them. Review and generalise. |
| Is the risk removed? | No, it is reduced. To remove it, you need a local AI. |
Every limit of Shield is on the Limits and compliance page. This post is not legal advice. It describes a tool.
Sources
- RAGbase Shield README (README.md and README.fr.md).
- Shield privacy model (docs/PRIVACY_MODEL.md) and French compliance sheet (docs/CONFORMITE.fr.md).
- GDPR, article 4(5) and recital 26: https://eur-lex.europa.eu/eli/reg/2016/679/oj
- EDPB, Guidelines 01/2025 on pseudonymisation