Summarize opposing filings with AI without breaching secrecy

Can you give the other side's filings to an AI?

Not in clear text. Opposing submissions, the list of exhibits and the documents disclosed describe your client, the facts and each side's strategy; once received, they are part of your file. In France, Article 66-5 of the Act of December 31, 1971 places under professional secrecy consultations, correspondence, interview notes “and, more generally, all the documents in the file” (our translation). The national bar rules (Règlement intérieur national, RIN) add “the names of clients and the lawyer's diary” (Art. 2.2), and describe secrecy as “general, absolute and unlimited in time” (Art. 2.1). Quotations from French texts in this article are our translations.

A prompt sent to an online generative AI transmits data to the company that runs it, which stores and processes it (CNB guide, p. 14). The practical guide of the Conseil national des barreaux, “La déontologie et l'intelligence artificielle” (April 2026), draws a precise rule from this: a lawyer may enter neither the client's name, nor the strategy of the case, nor attach “any file containing all or part of a case file and containing the client's name” (p. 15).

The same guide gives the way out. Asked how to have an AI “produce a summary”, it answers with pseudonymization or anonymization (p. 16), and it lists “prior anonymization or pseudonymization of the data” among the only ways to remove sovereignty risks (p. 26). Our reading of the CNB guide, point by point covers the rest.

One more point, specific to opposing filings: you do not own the rights to them. The guide asks lawyers to check that the tool's terms do not allow the provider to appropriate or reuse such material, and expressly mentions “opposing submissions” (p. 29). A business plan that does not train on your data is therefore preferable; see Claude, ChatGPT, Mistral or Copilot for a client file.

When is AI genuinely useful on the other side's file?

On reading and sorting tasks, where the reasoning does not depend on who the parties are. On substantive choices, it remains an aid to be checked.

Two uses stay out of scope: case strategy, which the CNB guide excludes from any prompt (p. 15), and case law: the guide lists several 2025 and 2026 French decisions dealing with untraceable authorities produced by AI (pp. 28-29). Any decision the AI cites is checked at the source, every time.

Why process the whole file, not just the opposing brief?

Because the AI only reasons correctly if each person keeps the same tag across every document. If the opposing company becomes [ORGANISATION_1] in its brief, [ORGANISATION_3] in yours and [ORGANISATION_2] in the contract filed as an exhibit, the AI sees three companies: the timeline blurs, and a “contradiction” may be nothing more than a masking artifact. Processed together, the documents share one mapping table, and the variants of a name (“Mrs Durand”, “DURAND Camille”, “Camille Durand”) are linked to the same person when there is no ambiguity.

The whole file does not mean sending everything. The CNB guide asks lawyers to limit what they share to “the data strictly necessary to perform the task” (p. 16). Good practice has two parts: pseudonymize everything, for consistent tags, then give the AI only the copies the question needs. For a summary, the last two rounds of briefs and the exhibit list are often enough; for a timeline, add the dated exhibits. A large file may exceed what a tool reads at once: work in batches (briefs, then exhibits) with the same instructions, and ask the AI to list anything it could not read.

What should be masked in a litigation file, and why?

Everything that makes the case recognizable, not just the names. The CNB guide targets “any specific detail of a case that could identify a person or a company” (p. 15) and warns about identification by cross-referencing, citing amounts and internal project names (p. 20).

Keep what the AI needs to reason: the statutes and published case law the parties rely on, the sequence of facts, the roles (“the landlord”, “the insurer”), durations and time limits. A simple test before sending: would a lawyer at the same bar recognize the case from the copy? Our article on re-identification risk covers these cross-references, and the guide to anonymizing legal briefs before AI lists the areas people forget (headers, exhibit list, metadata).

How to summarize opposing filings with AI, step by step

1. Gather the file. Your latest briefs, the other side's, the exhibit lists and the useful exhibits, in one folder on your computer. Scans go through local OCR.
2. Pseudonymize it all at once. In AnonymFiles, drop the whole folder, subfolders included: names, organizations, addresses, emails, phone numbers, IBANs, private dates and ID numbers are replaced by stable tags shared by every document.
3. Review side by side and make the tags meaningful. Each copy appears next to its original. Mask the docket number, amounts and references with one click: the rule applies to the whole file. Rename the key tags, for example [PERSONNE_CLIENTE] (client) and [ORGANISATION_ADVERSE] (opponent): the AI follows who argues what, and the name stays in brackets for reinjection. Never put initials or part of a name in a tag.
4. Validate, then export for the AI. Nothing leaves without your validation, document by document. The export holds the tagged copies, never the mapping table; the “Copier la consigne IA” (copy the AI instruction) button copies the text to paste at the top of the conversation.
5. Ask one question at a time. Paste the instructions, attach only the copies you need, then use the prompts below.
6. Check every statement against the original. Page, exhibit, date: the checklist is below.
7. Reinject the real names on your computer. Paste the answer into “Réinjecter un résultat” (re-inject a result): a preview shows the text with the real names, which you save as text, Word or PDF.

The full workflow, with screenshots, is on how AnonymFiles works for a law firm; other situations are gathered in the use cases.

Which prompts work for analyzing opposing submissions?

Five ready-to-copy requests, pasted after the instructions. Adapt the tags to your export: here, [PERSONNE_CLIENTE] is your client and [ORGANISATION_ADVERSE] the opposing party. The app's tags are in French; the AI handles them fine.

The note comes back with tags; reinjection puts the real names back, on your computer. More examples in ChatGPT and lawyers: the method.

What does the AI give you, and what must you check?

A first structured read, with tags: a working tool, not a document to file. The lawyer must “reread, check and validate every generated content before using it or sending it to a client or a court” (CNB guide, p. 28). The CNIL, the French data protection authority, also warns against excessive trust in generative AI output “without appropriate verification”.

How do you put the real names back into the summary?

On your computer, never by asking the AI. In AnonymFiles, “Réinjecter un résultat” accepts the answer pasted as is, or a .txt, .md, .docx or .pdf file. The [DOSSIER_…] reference finds the right mapping, even after a restart, without reopening the original file. The replacement uses no model and no connection. The mapping table is encrypted (AES-256-GCM) and never leaves the computer: see the mapping table, what it holds and how to protect it. Limits: each tag comes back in one form of the name, grammar is not adjusted, and a tag altered by the AI is flagged, never guessed. Read the preview.

Which rules apply to lawyers and in-house counsel?

• The French lawyer's professional secrecy. Article 66-5 of the Act of December 31, 1971 and RIN Articles 2.1 and 2.2. Correspondence between lawyers is also “confidential by nature” (RIN, Art. 3.1), except letters marked “officielle”: a letter from opposing counsel in the file is pseudonymized like the rest.
• The GDPR. Parties, witnesses and counsel are individuals: their data must be “limited to what is necessary” (Art. 5). Pseudonymized, it remains personal data (Recital 26), and the mapping table must be “kept separately” (Art. 4(5)). See ChatGPT and GDPR.
• In-house counsel in litigation. Article 66-5 covers lawyers (avocats), but company information may be protected as a trade secret, which requires “reasonable protective measures” (French Commercial Code, Art. L. 151-1): handing a file in clear text to an online service can weaken them. See in-house counsel and DPOs.

What AnonymFiles does, and what it does not

AnonymFiles, French document anonymization software, is a desktop app coming soon for Apple silicon Macs and Windows. It prepares the copy; it talks to no AI. It reads text and scanned PDFs (local OCR), Word files, emails, spreadsheets, images and text files. A person keeps the same tag across the whole file, every document is reviewed side by side before export, and a text PDF comes out with the same layout, each masked passage becoming a labeled black box whose text is actually removed.

Its limits: amounts are masked by hand, no detection is perfect (hence the mandatory review). It runs with no account and no telemetry. Pricing: €290 excl. VAT per seat per year, €240 from 3 seats, quote above 20 seats, with a 14-day trial (see pricing). For the firm: pseudonymization software for lawyers.