ChatGPT in a law firm: uses, limits, professional secrecy

Can a lawyer use ChatGPT? Yes, for a precise part of the work, and under conditions that are anything but details. The line runs between two uses that have nothing in common: the tool that helps you write, legitimate and effective, and the tool you let state the law or touch client files, where the ethical and judicial trouble begins.

This guide draws that line, without demonising and without selling a dream.

What ChatGPT does well in a firm

On writing work, the tool is a real assistant: rewording a letter, structuring an opinion from your own analysis, summarising a document you provide, explaining a notion to a client, preparing an outline for a hearing or a meeting. What these uses share: the substance comes from you, the tool shapes it, and you review it.

Used this way, it renders the same service as a fast, tireless junior you would never let sign anything.

Where it becomes dangerous: stating the law

A generative model produces plausible text; it does not consult a case-law database. Ask it for decisions on a specific point and it can invent some, with perfectly formed citations. The textbook case is well known: in 2023, a New York lawyer was sanctioned by a federal judge for filing submissions citing case law fabricated by ChatGPT. Nothing comparable has happened before a French court yet, and the rule that keeps it that way fits in one line: no reference leaves the firm unverified against official sources.

Legal research remains the domain of documentary databases and authentic sources; the generative model can help phrase the question, never certify the answer.

Professional secrecy, the wall that does not move

Article 66-5 of the law of 31 December 1971 covers opinions and correspondence between lawyer and client with professional secrecy, in all matters. Pasting the facts of a case into a consumer assistant means entrusting those elements to a third party, on infrastructure you do not control, under a jurisdiction that is not yours.

Business plans improve the contract: training excluded, confidentiality commitments, European data residency. They do not change the essential point: the operator remains subject to US law, and professional secrecy is not a negotiable clause, it is an obligation that follows you. The Conseil national des barreaux has taken up the subject and publishes its guidelines on generative AI; the profession is not discovering the problem, it is framing it.

Genuinely anonymising a case before submitting it is harder than it looks: dates, amounts and the sequence of facts are often enough to identify it. The simple rule: anything that would let a colleague recognise the case does not enter a pooled tool.

The rules of clean usage

1. Never a personal account for professional use; if the firm allows the tool, the firm provides the account and the settings.
2. Never identifiable facts or exhibits in a pooled tool, whatever the plan.
3. Every legal reference is verified against official sources before being used.
4. Whatever commits the firm is reviewed and signed by a lawyer; a model's output is a draft.
5. A written charter, even one page, applying to partners as well as associates and interns; our guide to governing usage at work gives the outline.

Beyond the assistant: agents on your files

The real potential for a firm is not a better chat window: it is AI that works on your files, at your place. An agent that reads the exhibits of a case and answers by citing the exact passage, that builds a chronology from your documents, that keeps a log of everything it did: the value comes from access to your material, and that access is only acceptable if the infrastructure is yours, in France, isolated, audited.

That is the Workspace model: your agents, on your data and your tools, on models you control, with one principle that does not vary: the agent prepares and cites, the lawyer decides and signs. Our approach to regulated professions starts there.

Frequently asked questions

Is a lawyer allowed to use ChatGPT?

Yes, no law forbids it. Professional ethics frame the usage: secrecy, competence, full responsibility for whatever leaves the firm. The tool creates no new duty; it makes it easier to breach the existing ones.

Can ChatGPT replace a lawyer?

No. It moves the line of the work: formatting, summarising and preparation can be delegated; analysis, strategy, counsel and signature stay with the lawyer. A firm that delegates preparation gains time; a firm that delegates judgement stakes its liability on plausible text.

Does ChatGPT know French law?

It has knowledge from its training: real, but with no guarantee of version, date or source. It can explain a notion correctly and be wrong about the current state of the law with the same confidence. For the state of the law, only official sources and documentary databases count.

Is there a French, sovereign legal AI?

The established legal publishers offer specialised AI for research and analysis, hosted on their own terms: ask them the jurisdiction and reversibility questions that apply. The other path is the one we defend: sovereign agents working on the firm's files, citing their sources, on French infrastructure you control. The two can coexist; let's discuss your case.