California in September, the French national bar council in April. Both address what is handed to a system, and what is signed afterwards.
On 30 September, California enacted Senate Bill 574, the first American statute devoted to the use of generative artificial intelligence by attorneys. It takes effect on 1 January 2027. An attorney may no longer delegate legal judgement, nor enter confidential or identifying information into a system whose access is not restricted, nor sign a citation they have not personally verified.
Six months earlier, in April, the Rules and Practice Commission of the Conseil national des barreaux issued its guide on professional ethics and artificial intelligence. It recommends systematically pseudonymising or anonymising data before any use of a tool, limiting the context transmitted, and comparing every text produced with the official decision.
A statute and a professional guide do not carry the same weight: one sanctions, the other translates into practice obligations that already existed. Bringing them together remains instructive, because they cut the problem the same way: what goes into the system, and what comes out under a signature.
The commentary in circulation keeps only one of the two, and often gets the first one wrong. The Californian statute prescribes no general anonymisation: it restricts access and requires verification. These are not the same obligations, and they do not call for the same measures.
In both systems the obligation rests on the professional, not on the supplier, and it changes nature: it is no longer enough to be diligent, it has to be demonstrable. The sanctions handed down bear on the signature, not on the tool.
Reading this month
The two obligations now framing AI
What each text provides, the sanctions already handed down, and five questions to ask first.
You are faster, so it should cost less?
The objection taken seriously, and the exact point where it turns around.
Case law searched from the matter
The count of the back-and-forth between the tool and the matter, and what it costs.
The full corpus : max-legal.ai/blog
Anonymise or pseudonymise
Anonymisation is irreversible: individuals can no longer be identified, by anyone, ever. Pseudonymisation replaces identifying elements with substitutes and keeps the means of reversing the process. In legal work it is always pseudonymisation, since the final deliverable must carry the real names. What has to be protected, then, is that return.
It is the distinction both texts of the month rest on: removing the names does not mean the same thing if the return remains possible.
The full definition →Worth noting
Approving content, authorising an action
On 9 October, Anthropic published an account of unintended actions by its models, with no customer data involved. A model told to stop short of submitting a form submitted it anyway. What the cases have in common is not the error but the persistence: faced with a restriction, the system looks for a way around it rather than stopping.
When a system prepares an email, fills in a form or edits a shared document, how far does it go on its own? Approving content and authorising an action are two distinct gestures, and the second does not follow from the first. MAX sends nothing and submits nothing: it produces a deliverable, and sending remains your gesture. Authorisation is not delegated to it because it was never entrusted to it.
Due diligence. You hand Max the matter with no further instruction, and you get the due diligence back.
It does not wait to be told what is wanted: it recognises what the matter calls for, sets the scope and produces the red flag report, the Q&A and the list of documents reviewed. It has become one of Max's own skills.
And when you do have an instruction, it prevails. A precise question, a point to check, three documents to pull out: Max answers what you ask, on the same material.
Not yet subscribed and would like to
Receive the next one
Once a month, at the start of the month.
Four minutes to read.