News in Practice
SB 574 and the CNB guide: the two obligations now framing the use of AI in legal work
A Californian statute and a French professional guide, six months apart, set the same requirements on two moments: what is handed to an artificial intelligence, and what is signed on the way out.
For lawyers, in-house counsel and data protection officers who have to settle a practice. What each text provides, the sanctions already handed down, and what now has to be demonstrable.
Within six months, two authorities from two different legal systems have written the same thing about the use of generative artificial intelligence by legal professionals. One through statute, the other through a professional guide.
Both address two distinct moments of the same work: what enters the system, and what comes out of it and ends up under a signature. The commentary in circulation generally keeps only one of the two.
What California passed
Senate Bill 574, chapter 858 of the 2026 statutes, was signed on September 30, 2026 and takes effect on January 1, 2027. It is the first American statute devoted to the use of generative artificial intelligence by attorneys.
The new section 6068.1 of the Business and Professions Code sets out four obligations. The attorney may not delegate the practice of law to a generative artificial intelligence, and remains responsible for the legal judgement underpinning the work. The attorney may not enter confidential, identifying or non-public information into a generative artificial intelligence system unless access to that system is restricted to the attorney and to authorised persons bound by confidentiality. The attorney must take reasonable steps to verify the accuracy of outputs, including every citation to case law and to statute, and must correct erroneous or fabricated results. And the attorney must disclose the use made of artificial intelligence to the court for covered filings.
Section 128.7 of the Code of Civil Procedure, amended by the same statute, adds a direct rule: a filing may contain no citation that the responsible attorney has not personally verified, whether or not it came from an artificial intelligence.
One point is frequently misreported: the statute does not prescribe general anonymisation. It restricts access and requires verification. These are not the same obligations, and they do not call for the same measures.
The sanctions, and who carries them
A failure to verify citations falls under the sanctions regime of section 128.7, which existed before this statute and now applies to it.
Proceedings open either on a party's motion or on the court's own initiative, and allow twenty-one days to withdraw or correct the filing before any sanction. Sanctions are limited to what suffices to deter, and may take three forms: non-monetary directives, a penalty paid into court, or the reasonable fees and costs of the party that brought the motion.
Two points matter more than the scale. No monetary sanction may be imposed on a represented party for a failure to verify citations: the burden falls on the attorneys. And absent exceptional circumstances, the firm is jointly responsible. The obligation is therefore not individual in the sense of stopping at the signatory.
The State Bar of California retains its own disciplinary jurisdiction, which operates independently of this regime.
The decisions that preceded the statute
The statute did not arrive in a vacuum. American courts had already begun to sanction, and that is what explains its drafting.
On June 3, 2026, in LNU v. Blanche, the United States Court of Appeals for the Ninth Circuit issued a published, and therefore precedential, order against two attorneys. Their briefs, filed in immigration appeals, contained non-existent authorities, citations attributed to the wrong decision, and seriously inaccurate characterisations of real decisions. The sanctions covered a sum of money, a six-month suspension from practice before that court, and an obligation on the firm to disclose, for two years, its use of artificial intelligence in its briefs, naming the tool and certifying that the signing attorney personally verified the citations.
The court was careful to state what it was sanctioning: the signature and the filing, not the recourse to artificial intelligence.
In April 2026, the State Bar of California had announced disciplinary proceedings against several attorneys over false content produced by an artificial intelligence and filed in court. One of them agreed to a suspension of their licence.
One detail of these matters deserves attention, because it shifts the nature of the check required. The complaints concerned not only invented decisions, but also citations attributed to the wrong decision and real decisions whose meaning had been distorted. Verifying that a reference exists is therefore not enough.
What the French bar council recommends
In France, there is no equivalent statute. The framework comes from elsewhere, and it is five months older.
The guide La déontologie et l'intelligence artificielle, first edition of April 2026, drawn up by the Rules and Practice Commission of the Conseil national des barreaux, recommends systematically pseudonymising or anonymising data before any use of a generative artificial intelligence tool, and limiting the context transmitted to what is necessary. It warns against re-identification by cross-referencing, that is, the possibility of recognising a person from information available elsewhere, and notes that pseudonymisation is not workable for every use.
On the second moment it is equally clear. The professional must re-read, check and validate every generated item of content before using it, and compare the text produced with the official decision. The guide lists French decisions already handed down, including one from the administrative court of Orléans, requiring that references found by whatever means be verified in future.
A professional guide is not a statute and creates no new obligation. It translates into practice obligations that already existed, legal professional privilege under article 66-5 of the law of December 31, 1971 and article 2 of the national internal rules.
Two texts, two moments
The two sets are not interchangeable. One is a statute carrying sanctions, the other a guide resting on pre-existing rules. Bringing them together remains instructive, because they cut the problem the same way.
The first moment is entry. What information is communicated to the system, and to whom does it become accessible. The second is exit. What the professional signs, and what is thereby engaged.
In both cases, and in both systems, the obligation rests on the professional and not on the supplier. And in both cases its nature changes: it is no longer enough to be diligent, it has to be demonstrable.
What goes in: removing names is not enough
The quick reading leads to a simple practice: replace the names, then send. It is useful and insufficient, for a reason that goes to the definition of privilege.
Legal professional privilege does not attach to identities, it attaches to the matter. Opinions, correspondence and case documents are covered. A litigation strategy, a draft transaction or a piece of know-how remain confidential once every name has been removed. Likewise, for the firm or legal department that retains the means of re-attribution, replacing identities does not take the processing outside the scope of personal data protection.
The symmetrical difficulty is less discussed and costs more in practice. A transformation applied without discernment damages the reasoning. Generalising a date to a year, or an amount to an order of magnitude, suits a statistical publication and makes it impossible to compute a deadline or assess a threshold. One and the same person may appear as agent in one document, director in another and signatory in a third: a transformation that merges those occurrences produces a false result, which the reviewer no longer has the means to detect.
Protection and fidelity are therefore assessed together. A transformation that protects well and destroys meaning does not protect, it moves the risk to the exit.
What comes out: verifying is not re-reading
Both texts use the same verb, and they denote the same operation, which is not the one common practice covers.
Re-reading a text means examining it in itself. Verifying a reference means leaving the text and going to the source. The French guide says it in plain terms, compare the text produced with the official decision. The Californian statute requires personal verification of every citation. In both cases the operation expected is a trip to the source, and it cannot be carried out on the document alone.
The decided cases show how far this goes. An invented reference is caught by looking for the decision. A citation attributed to the wrong decision, or a real decision whose meaning has been distorted, is caught only by reading the decision itself. The check therefore covers three points: the reference exists, it corresponds to the passage cited, and it says what it is made to say.
This is also why the check cannot remain with the last link in the chain. A professional who receives a document whose references have not already been confronted with their source has to do the work again in full, at the moment when there is least time for it.
Five questions before sending
The decision is not taken at the level of the document but at the level of the task. Five questions are enough to inform it.
Can the question be put in the abstract. To understand a general mechanism, stating the applicable law and a typical situation avoids communicating anything from the matter.
Which data are genuinely necessary. For each, state its function: to identify a role, establish a fact, compute a deadline, interpret a clause. An element with no function can be removed.
Is a return to identities required. If the result has to rejoin the matter with its parties, reversibility is necessary. For external circulation with no return to identities, the requirement is different and higher.
Does the content remain sensitive once the identifiers are dealt with. If so, the answer is not a further transformation, it is a reduction of the context transmitted or a different access environment.
Who will validate the result, and on what. Validating the preparation covers the context and the protections chosen, validating the result covers the facts, the references and the legal relevance. These are neither the same checks nor, often, the same people.
What the examination of a service covers
Since the obligation rests on the professional, it falls to them to know what the service they use does, and to be able to explain it.
The unit to examine is not a feature, it is the complete path of the information. The receipt of the request and the documents. The preparation, that is, converting a document or recognising its text, which often involves services before any legal processing. The successive operations and what passes between them. The return, which restores identities and therefore produces a deliverable that is identifying again. And what remains afterwards, documents, records and access.
A vocabulary is not a property. The words private, professional or secure describe nothing until they are specified by the scope of the data, the operations performed and the access opened at each stage. The same questions apply to the exit: against what were the references confronted, and what record remains of it.
We are publishing a white paper that develops the first of the two moments: the risk model, the mapping of the path, an evaluation protocol on fictional matters and a worksheet to be taken into a firm's own procedure. It is written to be applied to any legal artificial intelligence service.