By profession

AI for lawyers: what professional secrecy allows

The question is not whether AI belongs in a law firm, but which technical and organizational precautions reduce the risk enough for its use to hold up under professional secrecy. This page covers the rules that apply to lawyers in France.

Published on August 2, 2026 by the Moon AI team. Legal references checked at their source on that date.

The framework for a lawyer who uses AI

In France, a lawyer’s professional secrecy is a matter of public policy. Article 2.1 of the RIN, the national rules of the French bar, describes it as general, absolute and unlimited in time. Its legal basis is Article 66-5 of the law of December 31, 1971; a breach falls under Article 226-13 of the French Criminal Code, which provides for one year’s imprisonment and a €15,000 fine. Article 2.2 of the RIN specifies that secrecy covers consultations, correspondence, meeting notes and every document in the case file, whatever the medium, tangible or intangible. A prompt sent to an online service is a medium like any other. Two further regimes apply on top: the GDPR, since the firm is the data controller and the tool’s publisher its processor, and Regulation (EU) 2024/1689 on artificial intelligence (the AI Act), whose Article 4 has required, since February 2, 2025, a sufficient level of AI literacy from the people who use these systems. To date, the RIN contains no provision specific to artificial intelligence: the existing principles apply, unchanged.

Never pass to a generative AI a client’s name, confidential information the client has shared, elements of the case strategy, or a file containing all or part of the case file. source (in French)
Disclosing secret information by a person who holds it by status or profession is punishable by one year’s imprisonment and a €15,000 fine. source (in French)
A lawyer must ensure that secrecy is respected by anyone who works with them in their professional activity, and answers for any breach committed in this way: choosing a service provider engages the lawyer’s responsibility. source (in French)
A firm may only use processors that provide sufficient guarantees, and the processing must be governed by a written contract setting out its subject matter, its duration and documented instructions. source (in French)
Since February 2, 2025, Article 4 of Regulation (EU) 2024/1689 has required a sufficient level of AI literacy among the staff who use these systems. source (in French)

What AI does in practice

Each case lists the personal data it involves, and therefore what Moon Blur removes before sending to the model.

Researching case law and framing a legal question

Explore an applicable legal regime, surface possible legal characterizations, test a line of reasoning before going into the databases. The CNB, France’s national bar council, is clear: a generative AI system is not a search engine. Every reference it produces is checked in an authoritative database, including the number, the date and the holding of the decision.

Data involved: The legal question can be asked without the case file. If a document is attached, Moon Blur removes addresses, dates of birth and identifiers before sending to the model; names are masked only if they are in your terms to mask (from Constellation).

Summarizing a large case file

Rebuild a timeline across several hundred pages, spot contradictions between documents, isolate the points actually in dispute. The time saved is directly verifiable, since the source is in front of you: each statement in the summary can be checked against the document it cites.

Data involved: Names of the parties, addresses, social security numbers, IBANs, medical data, identity documents. Moon Blur replaces the addresses, numbers and identifiers it recognizes with placeholders, then restores them in the answer; names are masked only if they are in your terms to mask.

Analyzing and comparing contracts

Compare two versions of a draft, bring out the differences clause by clause, spot provisions missing from a reference template, extract deadlines and conditions precedent from a set of contracts. The result is a starting point for review, never a conclusion.

Data involved: Company names, SIREN numbers (French company identifiers), bank details, amounts, names and contact details of signatories. Moon Blur masks SIREN numbers, bank details and the signatories’ contact details before sending to the model; company names, amounts and names stay in plain text unless they are in your terms to mask.

First draft of a document: submissions, legal opinion, letter

Get past the blank page, produce a structure, rephrase a passage that is too technical for a client who is not a lawyer. The CNB lists drafting among the relevant uses. But the case strategy is never sent: it stays in your head and goes back in when you review.

Data involved: Contact details and identification numbers of the parties, removed by Moon Blur; names and identifying facts stay in plain text, except names added to your terms to mask. Strategy cannot be masked: it simply must not appear in the prompt.

Preparing oral argument and testing it against the other side

Have the tool play opposing counsel on a depersonalized argument, list the objections you can expect, test the ranking of your arguments and the order in which you present them. The exercise works on an abstract outline, which makes it compatible with secrecy and, incidentally, more useful.

Data involved: No names are needed. If documents are attached for context, Moon Blur removes contact details and identification numbers before sending (names only if they are in your terms to mask).

Translating documents and correspondence

Translate a contract, a foreign judgment or an exchange in a foreign language to understand it quickly, before ordering a sworn translation when the procedure requires one. Useful in international business law, immigration law and cross-border enforcement.

Data involved: Names, addresses, passport or residence permit numbers, bank details. Moon Blur removes addresses, passport or residence permit numbers and bank details before sending and restores them in the translation it returns; names are masked only if they are in your terms to mask.

Eight questions to ask about a tool

This checklist applies to any tool on the market. The last column gives Moon AI’s answer, including when it is no.

Checklist for evaluating an AI tool for lawyers
Criterion	Why it matters here	Moon AI
Where the firm’s data is hosted	The CNB puts this point first among its sovereignty criteria. A document covered by secrecy and stored outside the European Union escapes the European framework and your ability to have it deleted.	App and storage at OVHcloud, in Strasbourg and Roubaix. French publisher: Stellarr Studio SAS, SIREN 101 691 103.
Publisher’s nationality and extraterritorial exposure	The CNB recommends ruling out companies whose national law allows the authorities to requisition data stored on servers they own, wherever those servers are.	French publisher and French host. The third-party models it calls remain subject to the law of their own publishers: Moon Blur aims to reduce the exposure of the data it detects; it does not remove legal obligations or the risk tied to third-party providers.
No reuse of inputs for training	This is a CNB prerequisite, and the CNB immediately adds that a publisher’s commitment on this point is generally neither provable nor verifiable. A necessary criterion, never sufficient on its own.	No training on customer content. This commitment is declarative with us as anywhere else: Moon Blur exists so that you do not have to rely on it alone.
Reversible masking before sending to the model	The CNB names only two ways to eliminate the sovereignty risk for data entered into the tool: making the data fully non-identifying beforehand, or installing the model locally. Moon Blur is presented here as reversible masking: it guarantees neither full de-identification nor the elimination of the sovereignty risk.	Moon Blur detects more than 150 categories of data; names are not spotted on their own (personal terms to mask from Constellation). The data it detects is masked before sending, then restored on return; detection may be incomplete. No external certification attests to it.
Encryption, logging, access on the publisher’s side	The CNB asks for encryption in transit and at rest, access logging, and no way for the publisher’s employees to reach the firm’s data, or failing that, strict access rights.	Encrypted connections (HTTPS), logging, separation by workspace. No ISO 27001 certification and no HDS approval (France’s health data hosting certification) to date.
Processing agreement under Article 28 of the GDPR	The firm is the data controller. Without a written agreement setting out the subject matter, the duration, documented instructions and the list of subprocessors, a piece is missing from its own record of processing activities.	Data processing agreement and list of subprocessors published in the legal center. Keeping the firm’s record of processing activities remains your responsibility.
Verifiable sources in answers	The CNB requires sources throughout the generated content, with a direct link to the third party hosting the cited decision, while pointing out that this does not protect against hallucinations.	Sourced web search on the models that support it. Moon AI is not connected to any case-law database: checking remains manual.
Choice of model, legal bias and reversibility	The CNB notes that models trained mostly on Anglo-American law multiply common-law biases, and recommends favoring models from the civil-law tradition.	The catalog includes European models, Moon 6, our in-house model, and Moon Jurist 1, presented as specialized in French law, all interchangeable within the same conversation. None is trained on an independently verified civil-law corpus: the common-law bias noted by the CNB is not neutralized, only avoidable by choosing the model. Conversations and documents can be exported at any time.

When to rule out a tool

Objective criteria, verifiable in any publisher’s public documentation.

The firm’s data hosted outside the European Union, or hosted in the Union by an entity whose national law subjects it to extraterritorial disclosure orders.
No written processing agreement under Article 28 of the GDPR, or a refusal to disclose the list of subprocessors and the countries where they are based.
No reversible masking or pseudonymization between the interface and the model: the attachment goes to a third party as it is, with the names and identifiers it contains.
Terms of service that allow inputs to be reused to improve a model, including as an opt-out setting whose effect cannot be verified.
Legal answers without sources, or based on references produced by the model itself instead of pointing to an identifiable third-party database you can consult.

What Moon AI does not do

A page for professionals that concedes nothing is not credible to them.

Moon AI is not a case-law database. No built-in access to Légifrance (the official French legal database), Judilibre (the Court of Cassation’s open case-law search) or legal publishers’ collections: any reference it produces must be checked in an authoritative source before being cited.
No certification specific to the legal sector, no approval from the bar, no health data hosting. It is a general-purpose tool, not practice management software or a certified service.
Moon Blur removes the identifiers it recognizes (names only if they are in your terms to mask) but does not prevent identification by cross-referencing. An amount, an internal project name or a distinctive turn of phrase can still identify someone.
Some uses do not lend themselves to reversible masking, such as live notes of a client meeting. The CNB points this out: in those cases, no tool of this type is suitable.

€9.90 per seat per month, VAT included

Lune plan. More than 100 models included, Moon Blur on, hosted in France. 20% off each extra seat in the same workspace. No commitment. See all five plans.

Start for free

Frequently asked questions

Can I use generative AI without breaching professional secrecy?

Yes, but not by sending the case file. The CNB, France’s national bar council, sets the rule in its practical guide of September 2024: a lawyer must never pass data covered by professional secrecy to a generative AI, and may not enter in a prompt the client’s name, confidential information the client has shared or elements of the case strategy, nor attach a file containing all or part of the case file. In return, the CNB recommends a good practice: pseudonymizing the data before sending. Words matter, and here they are legal ones: pseudonymization is not irreversible. It does not take the information outside the scope of secrecy, and it does not stop it being personal data. It reduces the exposure surface, it transfers no responsibility, and judging what may leave the firm remains up to you.

Is the AI I use at the firm a high-risk system under the AI Act?

In principle, no. Annex III, point 8(a), of Regulation (EU) 2024/1689 classifies as high-risk the systems intended to be used by a judicial authority or on its behalf to assist it in researching and interpreting facts and the law, or in alternative dispute resolution. An assistant you use for your own advisory or litigation work does not fall into this category. Article 4 of the same regulation, on the other hand, has applied to you since February 2, 2025: it requires a sufficient level of AI literacy among the people who use these systems, which means training the firm’s lawyers and staff.

What does a lawyer risk by citing a decision invented by an AI?

The CNB recalls two precedents in its guide. A New York lawyer cited seventeen decisions that did not exist, having failed to check a result obtained by using a chatbot as a search engine. More insidiously, the CNB describes a case where the reference, the date and the number of a ruling were correct, but the summary covered an entirely different case. The risk lies in the gap between how plausible the text is and whether it is true, all the more so as these systems express themselves assertively. The CNB draws a simple rule from this: the lawyer becomes the one who checks and validates the generated output.

Should I tell my client that I use AI on their case?

To date, the RIN contains no provision specific to artificial intelligence, and the CNB states that its committees will rule on the impact of AI on the client relationship. What it recommends now concerns the firm’s internal organization: set out the authorized uses in an IT use policy, explain how and for what purposes AI is used to the whole firm, and train the staff. Toward the client, transparency is today a matter of the duty of loyalty more than of any written rule, and is usefully addressed in the fee agreement.

Is reversible masking enough to protect me?

No, and the CNB says so clearly. Identification remains possible by cross-referencing, even after pseudonymization: an amount, an internal project name, a distinctive way of drafting a contract can be enough to trace back to the client. Some uses do not lend themselves to it at all, such as live notes of a meeting. And elements of the case strategy cannot be masked, since their very content is what is confidential. A reversible masking mechanism reduces the exposure surface; it replaces neither your judgment nor your obligations.

How much does Moon AI cost, and is there a commitment?

Public prices range from €9.90 to €249.90 per seat per month, VAT included, depending on the level of use, with 20% off extra seats in the same workspace. There is a Free plan with no credit card, and no minimum term. Details are on the pricing page, and the contractual documents, including the data processing agreement, are in the legal center.

Sources

References checked on August 2, 2026. This page describes the law that applies in France as it stood on that date and is not legal advice. If you are unsure about a specific situation, consult your professional body or an adviser. Spotted an error? Write to contact@stellarrstudio.com.