I have watched two jurisdictions answer the same worry in different ways. California is close to putting its AI rules for lawyers into statute, where England and Wales leaves the ground to the SRA Code, to guidance and to what a judge says when a filing goes wrong. Senate Bill 574 sits on the Assembly floor in Sacramento, amended there on 21 August 2026, and 31 August is the last day for each house to pass bills in this session. Nothing in it will ever bind your firm. It repays a reading all the same.
The Senate carried it 39 votes to nil in January and the Assembly committees carried it 15 to nil and 11 to nil. Not one legislator has voted against it at any stage, which tells you how settled the worry has become in the place where most of these tools are built.
The four duties the bill would create
Section 6068.1, which the bill adds to the Business and Professions Code, says that an attorney shall not delegate the practice of law to generative artificial intelligence. Three requirements then follow for any attorney who uses it in practice.
The first governs what you put in. You must not enter confidential, personal identifying or other nonpublic information into a generative AI system unless access to what you input is restricted to you and to people you authorise who are themselves under obligations to protect it. The bill spells out what personal identifying information covers, from dates of birth and account numbers to the addresses and telephone numbers of parties, witnesses and court personnel.
The second is verification. An attorney must take reasonable steps to check the accuracy of outputs, all case and statutory citations included, and to correct any erroneous or hallucinated output. A parallel amendment to section 128.7 of the Code of Civil Procedure bars any paper filed in any court from carrying a citation the responsible attorney has not personally verified, whether or not a machine supplied it. The sanctions regime around it holds a firm jointly responsible for the failures of its partners, associates and employees.
The third is disclosure. An attorney would have to tell the court that generative AI was used, for every document submitted to it. Arbitrators draw a clause of their own, barring them from delegating any part of the decision-making process and from relying on machine-generated material outside the record without telling the parties beforehand.
Which of these you already owe
Set that list beside your own position and the overlap runs close to complete. The bar on delegating the practice of law restates what the SRA warning notice of 17 August puts in plainer words, that the solicitor who signs the work answers for it. The confidentiality duty covers ground the Code already covers, and checking authorities before they reach a court is settled here by decisions that have cost lawyers wasted costs orders and referrals to their regulators.
Disclosure is where the two part. England and Wales asks for no routine declaration that a machine helped produce a document. The judicial guidance rests responsibility on the person who signs, and the Civil Justice Council has been working out where a court needs to be told. California would settle the question by statute, for every document, in every case.
One line is worth lifting whatever becomes of the bill. Its confidentiality test does not ask whether a tool is secure, a question no supplier answers badly. It asks whether access to what you put in is restricted to you and to people you authorise who are under obligations to protect it. Most procurement forms never reach that far, and a supplier can answer it in writing or decline to.
Take the four duties as a single page and read your AI use policy against them, marking what your policy leaves untouched. Then put the access question to your supplier before the next renewal. Most of the tools on an English desk are built where this bill would bite, so the controls it forces will arrive inside the product whether or not the statute reaches you.
The bill is published as Senate Bill 574 on the California Legislature's own site, which is open to read and asks nothing of you.
If your AI use policy has not been read against what the regulator and the courts now expect, we do that in a morning: start with a conversation.
