ChatGPT for lawyers: can you use it without breaching professional secrecy?

Can a lawyer use ChatGPT?

Yes, but never on documents in plain text. Professional secrecy covers every document in the file, and everything you type or attach in ChatGPT is sent to OpenAI. The Conseil national des barreaux (CNB, the French national bar council) therefore recommends that lawyers “systematically pseudonymize or anonymize data before any use of a generative AI tool” (our translation).

In practice, ChatGPT can summarize a file, build a chronology or proofread a draft from a copy in which the client, the opposing party and third parties are replaced by labels such as [PERSONNE_1]. It replaces neither a reliable case-law database nor your own analysis: everything it produces must be checked.

What do French ethics rules say about ChatGPT?

They do not forbid it, but they forbid putting into it anything covered by secrecy: in practice, any document that has not been pseudonymized.

Secrecy covers the whole file, down to the client's name

Article 66-5 of the Law of December 31, 1971 covers advice, correspondence, meeting notes “and, more generally, all documents in the file.” The Règlement intérieur national (RIN, the national internal rules of the bar) states that this secrecy is a matter of “public policy” and is “general, absolute and unlimited in time” (art. 2.1). It covers every medium, “tangible or intangible,” and extends to “the names of clients” (art. 2.2). Typing a client's name into ChatGPT already means sending it to a third party.

The lawyer is also answerable for their team. They must ensure that secrecy is respected by the firm's staff and by anyone working with them (art. 2.3): “They are answerable for any breaches of secrecy thus committed.” An assistant or associate who pastes a document in plain text into ChatGPT therefore engages the lawyer's liability. The CNB guide accordingly recommends an AI usage policy within the firm (p. 36): see our annotated AI policy template.

In-house counsel fall under a different regime: Law No. 2026-122 of February 23, 2026 creates confidentiality for their legal opinions, from a date set by decree. Article 66-5 and the RIN apply to lawyers (avocats).

What the CNB guide of April 2026 recommends

In March 2026, the CNB adopted a practical guide, “La déontologie et l'intelligence artificielle” (Ethics and artificial intelligence, 1st edition, April 2026). Its starting point is strict: never communicate data relating to files or clients to a generative AI. A request must therefore contain neither the client's name, whether an individual or a company, nor the confidential information they entrusted to you, nor the strategy of the matter, nor a file that contains any of these.

The guide does not close the door, though. For a summary, part of a written submission or a spelling check, it accepts one solution: pseudonymize or anonymize the data. It also asks lawyers to limit the context to what is strictly necessary and to beware of indirect identification through cross-referencing. The detail, recommendation by recommendation: the CNB guide on AI, point by point.

Pseudonymizing means replacing names with aliases while keeping what you need to restore them; anonymizing means making any identification impossible, irreversibly. To put the real names back into the AI's answer, pseudonymization is the right tool. The detail is in our article on the difference between anonymization and pseudonymization.

The criminal and disciplinary risk

Disclosing secret information by a person who holds it through their profession is punishable by one year of imprisonment and a €15,000 fine (Article 226-13 of the French Criminal Code, Code pénal). According to the CNB guide, sending case data to a generative AI may expose the lawyer to disciplinary or legal sanctions, and using generated content without checking it may engage their civil liability.

What risks does ChatGPT pose for a law firm?

Four: data leaving the firm, its reuse depending on the plan, the absence of any secrecy over the conversation, and plausible but sometimes made-up answers.

The data leaves the firm

The CNB guide describes the mechanism: every request sent to a generative AI transmits data to the company that runs it; that company stores and processes it, sometimes to improve its models. For ChatGPT, that company is OpenAI. Once sent, the text follows the provider's rules, no longer the firm's.

Reuse that depends on the plan

Depending on the plan, the text you send may be used to train OpenAI's models and is kept for a longer or shorter time: the detail, plan by plan, is in the section which ChatGPT plan for a law firm.

No secrecy for a conversation with an AI

OpenAI's CEO has acknowledged it himself. On Theo Von's podcast This Past Weekend (episode 599, late July 2025), Sam Altman pointed out that something confided to a therapist, a lawyer or a doctor is protected, “there's legal privilege for it,” whereas a conversation with ChatGPT is not: in a lawsuit, OpenAI could be required to produce it.

A US judge has drawn the consequences. In February 2026, in United States v. Heppner, Judge Jed Rakoff of the US District Court for the Southern District of New York held that a defendant's exchanges with Claude, Anthropic's consumer assistant, were protected neither by attorney-client privilege nor by the work-product doctrine. Three grounds: a chatbot is not a lawyer; the vendor's privacy policy allowed the data to be collected and disclosed, which ruled out any reasonable expectation of confidentiality; and the defendant had acted on his own initiative, without instructions from his lawyers.

This is US law, handed down by a trial court: it does not apply in France. Yet it illustrates what the CNB guide already says: a text given to an online AI is disclosed to a third party. And the case concerned the client, not their lawyer: a useful reminder for clients who “prepare” their case with a chatbot.

Plausible answers, sometimes made up

ChatGPT produces plausible text, not documentary research. OpenAI's terms of use for Europe say so themselves: answers are not always accurate, and they should not be relied on as a substitute for professional advice.

French courts have already noticed. The CNB guide cites a decision of the Périgueux judicial court of December 18, 2025 (No. 23/00452): the case-law references cited by the claimant “do not appear to correspond to published decisions” (our translation). It also cites a decision of the interim relief judge of the Bordeaux administrative court of appeal of February 26, 2026, noting that two Conseil d'État decisions relied on do not exist. Hence the guide's rule: all generated content is reread, checked and validated before being used or sent.

Which ChatGPT plan for a law firm?

A business plan, ChatGPT Business or Enterprise, with a signed data processing agreement (DPA): that is the minimum. But no plan replaces pseudonymization, which the CNB guide retains as the only safeguard together with a model installed at the firm. Here is what OpenAI's help and privacy pages state, as reread in October 2026:

Two clarifications: even after turning training off, giving a thumbs up or down on an answer makes the whole conversation usable for training; and not training is not the same as not receiving, since the text is sent to OpenAI and kept, at least for a while.

The CNB guide advises against relying on the terms and conditions alone: when a provider states that data will not be reused, that commitment “can generally be neither proven nor verified” (p. 26, our translation). It retains only two safeguards: “prior anonymization or pseudonymization of the data,” or a model installed on the firm's own systems. A business plan reduces the risk; it does not exempt you from pseudonymizing. The GDPR obligations of each plan are detailed in ChatGPT and GDPR: can you put client data into it?

ChatGPT or a legal AI: what does the CNB recommend?

For legal research, the CNB guide recommends the AI tools of legaltechs and major legal publishers, “whose databases are more reliable” (p. 25, our translation). ChatGPT remains useful for summarizing, rephrasing or proofreading, on pseudonymized copies.

To choose any tool, the guide proposes three families of criteria:

• Data sovereignty: hosting location, the provider's nationality, no reuse for training (p. 26).
• Security: encryption with a key held by the firm, deletion of requests, restricted access for the provider's staff, logging (pp. 26-27).
• Reliability: cited, verifiable sources, and a model suited to the applicable law (p. 27).

The “Grille de lecture” (assessment grid) published by the CNB's AI working group in June 2025 points the same way: to eliminate sovereignty risks, it retains only “complete prior anonymization of the data” or a model installed on the firm's own instances (p. 11, our translation). The detail, criterion by criterion: the CNB guide on AI, point by point.

What can you give ChatGPT, and what should you keep to yourself?

On a pseudonymized, proofread copy, ChatGPT saves time on sorting and formatting:

• Summarize a file: each party's position, the points of disagreement, what is missing.
• Build a chronology: facts, procedural steps and exhibits, in order. A masked date keeps its label; it will come back with the names.
• Draw up the table of claims: claims, legal grounds relied on, exhibits cited.
• Rephrase: clarify a note, adapt a letter for a non-lawyer client.
• Proofread a draft: typos, repetitions, ambiguous sentences, references to exhibits.

What not to do:

• Paste documents in plain text, even as a test: the client's name is already covered by secrecy.
• Hand it case-law research without checking: every reference is checked on Légifrance or a reliable database, and any summary of a judgment is compared with the official decision, as the CNB guide advises.
• Explain the strategy of the matter to it: the guide lists it among what a request may not contain.
• Have it write up a client meeting live: no pseudonymization is then possible, the guide notes.
• Leave the decision to it: the AI never replaces the lawyer, who remains responsible for their choices and advice.

How to use ChatGPT without breaching secrecy: the 5-step method

For a single document, our guide to anonymizing a document before sending it to ChatGPT details the steps. For a complete file, the logic is the same:

1. Pseudonymize the file. Work on a copy. Replace each name with a stable label, the same across every document: [PERSONNE_1] for the client, [PERSONNE_2] for the opposing party, [ORGANISATION_1] for the company involved. Also handle addresses, e-mails, phone numbers, IBANs and ID numbers. Keep the mapping table at the firm, never with the copy, and avoid online anonymization tools, which add a third party (see anonymizing a PDF online or offline).
2. Proofread the copy. Look for anything that still identifies the matter: court docket number (RG), court, hearing dates, amounts. The CNB guide adds internal project names and the way a contract is drafted, which can be enough to recognize a client. Remove PDF metadata and comments in Word files, and handle scanned documents separately.
3. Paste an instruction. At the start of the conversation, ask the AI to keep the labels intact. For example: Names have been replaced by labels in square brackets, such as [PERSONNE_1]. Keep them exactly as they are, without changing them or trying to find out who they refer to.
4. Work with the AI. One task per request, with the source document for each statement. Check every case-law reference at the source.
5. Put the real names back. On your computer, replace the labels in the answer with the real values, then proofread: the AI may have attributed a fact to the wrong label.

On a file of several dozen documents, steps 1, 2, 3 and 5 take time. AnonymFiles, French document anonymization software, is a desktop app coming soon for Apple silicon Macs and Windows, is designed for them. It reads text and scanned PDFs (local OCR), Word files (.docx), e-mails (.eml, .msg), spreadsheets (.xlsx, .csv), presentations (.pptx), images, .txt and .md, detects names of people, organizations, addresses, e-mails, French-format phone numbers, IBANs from nine European countries, private dates and ID numbers, and replaces them with stable labels across the whole file. Amounts and other passages are masked in one click, during the side-by-side review: no detection is perfect, and nothing is exported without your approval.

The export truly deletes the masked content, removes metadata, annotations, forms and links, and gives the files neutral names; the mapping table stays on the computer, never in the export. The app then provides the instruction to paste into the AI, and lets you put the real names back into its answer, as text, Word or PDF, on the computer. The copy of a Word document is a redacted .docx that keeps its layout, without comments or tracked changes. No account, no telemetry, no online license check: on Mac, the models are included, with no connection at all; on Windows, they are downloaded once, at the first processing run, without any document text. See the security and network scope page, the published detection benchmark and offline test, and our anonymization software for lawyers. To see how AnonymFiles compares with other tools: the comparison of anonymization software.

Which prompts to use on a pseudonymized copy?

Four short requests to adapt, once the instruction is pasted and the copies are proofread:

• Summarize a document: Summarize the attached document in ten lines: what [PERSONNE_1] is asking for, on what grounds, and how [ORGANISATION_1] responds. Give the page for each point.
• Build a chronology: Build the chronology of the facts as a table: date, fact, source document. Keep the labels as they are and flag contradictions between documents.
• Draw up the table of claims: Draw up a table of [ORGANISATION_1]'s claims in these submissions: claim, legal ground relied on, exhibits cited. Do not add any reference that is not in the text.
• Proofread a draft: Proofread this draft letter to [PERSONNE_1]: flag ambiguous sentences, repetitions and typos, without changing the meaning or the labels.

Requiring the source for each point speeds up checking. For a judgment, see also our guide to anonymizing a judgment or court documents.