Anonymize legal briefs before ChatGPT: the lawyer's checklist

Why anonymize legal briefs before ChatGPT?

Written submissions concentrate the most identifying parts of a file: the client's name and the opposing party's, the facts, the strategy. Under Article 66-5 of the Law of December 31, 1971, “all documents in the file” are covered by professional secrecy. The Règlement intérieur national (RIN, the national internal rules of the bar) also includes “the names of clients” (RIN, art. 2.2). Yet every request sent to a generative AI transmits data to the company that runs it (CNB guide, p. 14).

The practical guide of the Conseil national des barreaux (CNB, the French national bar council), “La déontologie et l'intelligence artificielle,” adopted in March 2026 (1st edition, April 2026), takes written submissions as its very example. To ask an AI to “draft part of the submissions,” it answers: “The solution lies in pseudonymizing or anonymizing the data.” (p. 16, our translation). Its summary of good practices is clear: “Systematically pseudonymize or anonymize data before any use of a generative AI tool.” (p. 18). Its other recommendations are detailed in the CNB guide on AI, point by point.

For written submissions, pseudonymization is the most practical. Each name becomes a stable label: [PERSONNE_1] for the client, [ORGANISATION_1] for the opposing company. The AI keeps track of who does what, and you keep what you need to restore the real names. The guide points out that it is not irreversible, “unlike anonymization” (p. 16). See the difference between anonymization and pseudonymization.

The opposing party's submissions are handled the same way: they also name your client. The guide specifies that the lawyer does not hold the rights to them. The lawyer must therefore check that the AI's terms of use do not allow the provider to appropriate or reuse them (p. 29).

What should you mask in a brief before ChatGPT? The checklist

For each item: where to find it in written submissions, and whether AnonymFiles spots it on its own or whether it is masked by hand, in one click, during review.

Detected automatically, to be checked during review:

• Parties and representatives, people and organizations: heading (“For…”, “Against…”), facts, discussion, operative part.
• Lawyers and opposing counsel, with their firms and contact details: appearance (“represented by”), signature, letters quoted.
• Witnesses: statements quoted, list of exhibits.
• Experts: appointment, expert report, observations.
• Company names: insurer, bank, employer, supplier mentioned. A trade name or product needs checking.
• Addresses: home, registered office, property or site in dispute.
• E-mails: lawyers' contact details, exchanges reproduced.
• Phone numbers in French format: contact details, calls or texts quoted. A foreign number needs checking.
• IBANs from nine European countries: payments, transfers, refunds claimed. Others need checking.
• Dates of birth, as private dates: the parties' civil status, in the heading. Place of birth and occupation need checking.
• ID numbers: companies' SIREN numbers, social security numbers.

To check specifically or mask by hand:

This list and this table describe the detection, not a guarantee: no automatic detection is perfect, and every document is reviewed side by side with the original. Watch for variants of the same name: “Mrs. Durand,” “DURAND” in the heading, “née Martin,” initials. All must get the same label, in every document.

Keep, on the other hand, what the AI needs to reason: the statutes and published case law you cite, the sequence of facts, their legal characterization. An earlier decision handed down between the same parties, however, is masked like the docket number.

Cross-referencing remains. The CNB guide's summary points it out: “Be alert to the risk of indirect identification by cross-referencing data.” (p. 18). It cites amounts and internal project names (p. 20). The docket number (RG), for its part, leads straight to the parties: it is the case's registration number in the general register of the court registry, which also records “the names of the parties” (Code of Civil Procedure, Art. 726). A simple test: would a colleague at the same bar recognize the case by reading the copy?

AnonymFiles, French document anonymization software, is a desktop app coming soon for Apple silicon Macs and Windows: it handles the automatic part of this list, on your computer. It reads text and scanned PDFs (local OCR), Word files (.docx), e-mails (.eml, .msg), spreadsheets (.xlsx, .csv), presentations (.pptx), images, .txt and .md. It spots names of people, organizations, addresses, e-mails, French-format phone numbers, IBANs from nine European countries, private dates and ID numbers. It replaces them with stable labels: the same person keeps the same label across every document. A passage selected during review, such as a docket number or an amount, is masked across the whole file, and nothing is exported without your approval. The mapping table stays on the computer, never in the export. The workflow is detailed, with fictitious submissions as an example, in how AnonymFiles works for a law firm. See also our anonymization software for lawyers, the published detection benchmark and the comparison of anonymization software.

What can ChatGPT do with anonymized briefs?

On a pseudonymized, proofread copy, ChatGPT helps sort, structure and proofread; it replaces neither case-law research nor strategy. A few requests, once the instruction is pasted:

• Summarize the opposing submissions: Summarize [ORGANISATION_1]'s submissions: claims, arguments, exhibits relied on, with the page of each item.
• Build a table of claims: Make a table of the claims: claim, legal basis relied on, exhibits cited, possible response. Do not add any reference that is not in the text.
• Build a chronology: Build the chronology of the facts: date, fact, exhibit. Flag contradictions between the submissions and the exhibits.
• Proofread a section: Proofread the section "On liability": ambiguous sentences, repetitions, missing references to exhibits, without changing the labels.

What not to ask it: to set out the strategy of the case, which the CNB guide lists among what a request may not contain (p. 15), or to cite case law without checking: the guide reports 2025 and 2026 decisions pointing to references that could not be found (pp. 28-29). More examples in ChatGPT for lawyers: the 5-step method.

Which areas do people forget? Headers, exhibit list, metadata

• Headers and footers. They often repeat, page after page, the docket number, the case title (“Durand v. SCI Les Tilleuls”) and the firm's contact details. In a Word file, AnonymFiles handles them along with the body text and footnotes.
• The list of exhibits. It often names the authors of statements and letters, and dates each exhibit: “Exhibit No. 7: statement by Mr. Paul Garnier.” Names are detected; invoice or contract numbers need checking, and are masked in one click.
• Comments and tracked changes. A draft brief attached as a .docx carries them along, with the file's properties. According to Microsoft, these traces can reveal “the names of people who worked on your document, comments from reviewers.” Yet a partner's comment may contain the strategy of the matter, which the CNB guide excludes from any request (p. 15). To clean the file yourself, see anonymizing a Word document without putting it online. In AnonymFiles, a Word file gives a .docx redacted in place, layout preserved, from which comments, tracked changes, images and properties are removed, plus a text copy for the AI.
• PDF metadata and the file name. “Conclusions Durand v3.pdf” goes along with the attachment. On export, AnonymFiles removes metadata, annotations, forms and links, and gives the copies neutral file names. By hand: removing PDF metadata.

How to anonymize the exhibits, scans included?

Exhibits follow the same rule as the submissions: same labels, same checks, and only those the AI needs. A contract filed as an exhibit is handled with our checklist to anonymize a contract before ChatGPT, also written for in-house counsel; an e-mail produced as an exhibit, with the guide to anonymize an e-mail (.msg, .eml).

The exhibits exchanged are often scans: handwritten statements, signed letters, official reports. AnonymFiles reads them through local OCR and erases the pixels of the detected areas; review them side by side, because a stamp or handwritten signature that the OCR cannot read is masked in one click. For your own submissions, start from the Word file or the PDF produced by your word processor: the copy keeps its layout.

For these documents, three options. Do not attach them, if the task does not need them: the CNB guide limits the context to the “data strictly necessary to carry out the task” (p. 16). Retype only the useful passage, then handle it like the rest. Or anonymize the scanned document by hand. The judgment under appeal is handled like the submissions if it is a text PDF: see anonymizing a judgment or procedural documents.

How to check the copy before sending it to ChatGPT?

Check the copy that will be sent, not the original, and in full:

1. Read the copy on its own. In AnonymFiles, the “copie seule” (copy only) view lets you reread each document in full, as the AI will receive it.
2. Search for your test words. Run a search (Ctrl+F or Cmd+F) for the client's name and the opposing party's, the city, the street, the docket number, a maiden name. No result should come up.
3. Test a PDF with copy and paste. If you redacted a PDF yourself, select a blacked-out area, copy it, then paste it into a text editor. If the text reappears, it was only covered: see why a black box leaves text readable.
4. Do the colleague test. Would a lawyer at the same bar recognize the case? If so, mask more: a date, a place, an occupation.

Which instruction to give the AI, and how to put the real names back?

At the start of the conversation, before the submissions, paste an instruction that protects the labels. For example:

Names and identifying data have been replaced by labels in square brackets, such as [PERSONNE_1] or [ORGANISATION_1]. Keep each label exactly as it is, without changing it, translating it or trying to find out who it refers to. Do not invent any new label.

In AnonymFiles, the “Copier la consigne IA” (copy the AI instruction) button provides these instructions after export. They also ask the AI to restate the file reference in its answer, in the form [DOSSIER_…]. To help the AI write, give a label the name of its role: in AnonymFiles, [PERSONNE_1] can become [PERSONNE_CLIENTE]. Never put real initials in a label.

Once the answer is in, the real names are put back on your computer, never in the conversation. By hand, it is a Find and Replace, label by label. In AnonymFiles, “Réinjecter un résultat” (re-inject a result) accepts the pasted answer or a .txt, .md, .docx or .pdf file, and shows a preview before saving. A Word file is restored in its layout, and grammar is not adjusted: check agreement.

The substantive check remains. The CNB guide asks lawyers to “reread, check and validate every generated content” before sending it to a court (p. 28). Every case-law reference is checked at the source, every fact compared with its exhibit. The AI may also have attributed a fact to the wrong label. For the tasks it can handle on a pseudonymized copy, see ChatGPT and the lawyer's professional secrecy.