Senate Bill 574, signed September 30, puts generative AI rules for California attorneys into state statute starting January 1, 2027. That gives firms about ten weeks to answer a practical question: our people already use AI, so what do we have to change?
The law firm alerts published since the signing do a good job of explaining what the statute says. Few of them get into the operational work it creates, and much of that work lands on whoever runs the firm’s technology. Whether a firm can lawfully put client information into a given AI tool now depends on how that tool handles data. Verification and court disclosure need a workflow and a record. This article covers the setup side. For legal interpretation, talk to your ethics counsel.
What Changed
SB 574 adds Section 6068.1 to the Business and Professions Code and amends Section 128.7 of the Code of Civil Procedure. Baker McKenzie describes it as making California the first state to regulate attorneys’ use of generative AI by statute. Much of it builds on duties the State Bar’s practical guidance already described, first issued in 2023 and updated in May 2026. The difference is that the core expectations are now written into law, and the citation rule comes with court sanctions.
The bill also covers arbitrators and alternative dispute resolution providers. Those provisions are outside the scope of this article.
The Requirements in Plain English
AI can assist, but it can’t practice law
The statute says an attorney “shall not delegate the practice of law to generative artificial intelligence.” It doesn’t define “delegate.” Holland & Knight reads it to allow AI help with research, drafting, document review, and correspondence, as long as the attorney remains responsible for the legal judgment behind the work.
Confidential information only goes into restricted systems
Attorneys may not enter confidential, personal identifying, or other nonpublic information into a generative AI system unless access to what they enter is restricted to the attorney and people the attorney has authorized who are bound to keep it confidential.
The statute’s list of personal identifying information is broad: driver’s license and Social Security numbers, dates of birth, addresses and phone numbers of parties, victims, witnesses, and court personnel, medical or psychiatric information, financial information, account numbers, and anything sealed or confidential by court rule or statute. Baker McKenzie notes there’s no general exception for information that is already public, such as a witness’s address.
Check what the AI produces, and fix it
Attorneys must take reasonable steps to verify the accuracy of AI output, “including, but not limited to, the accuracy of all case and statutory citations,” and correct any erroneous or hallucinated output in material they use. This applies to AI-assisted work generally, whether or not it ever reaches a court.
Every citation in a court filing must be personally verified
A brief, pleading, motion, or other paper filed in any California court can’t contain citations the responsible attorney hasn’t personally verified. This applies to every citation, whether it came from AI, an associate, a paralegal, or a research database. Violations fall under Section 128.7’s sanctions framework, and absent exceptional circumstances the law firm is jointly responsible for violations by its partners, associates, and employees. Discovery papers are excluded under subdivision (g).
Disclose AI use to the court
Attorneys who use generative AI must disclose that use to the court “for all documents submitted to the court,” and consider disclosure when AI is used to create content provided to the public. The statute doesn’t say what form the disclosure takes or how much AI involvement triggers it. Holland & Knight suggests document-specific disclosure whenever AI was used to prepare a filing as the more prudent course until that’s clarified.
The Confidentiality Rule Is a Technology Decision
The confidentiality requirement turns on a technical fact: who can access what an attorney types into the tool. The statute doesn’t define “restricted,” and it doesn’t name any approved products. Baker McKenzie makes the practical point directly: an enterprise AI subscription alone does not establish compliance. Attorneys still need to confirm who can access the information and what confidentiality obligations apply.
That means each AI tool the firm allows for client work needs answers to the same set of questions. None of them has a universal answer. The answers depend on the vendor, the plan, and how the firm has configured it.
Personal plans are governed by the vendor’s consumer terms. Business and enterprise plans usually come with a contract, admin controls, and data protection terms. Read what the contract actually says about access and confidentiality.
Check the default for your plan and whether an administrator can lock the setting off for every user, so it doesn’t depend on each person remembering.
Many services keep conversations for a period and allow limited human review, for example for abuse monitoring. Find out the retention period, whether you can shorten it, and what exceptions the terms allow.
Shared workspaces are useful, but a project containing one client’s matter shouldn’t be visible to everyone in the firm. Check default sharing settings and who can invite others.
Single sign-on through the firm’s identity system means accounts are created and removed with employment, and nobody keeps client data in a tool after they leave.
Copilot works inside your existing Microsoft 365 tenant and can surface anything a user already has permission to open. If matter files are shared too broadly in SharePoint or Teams, Copilot will expose that. Some features, such as web search, handle data differently from the rest of Copilot. Review both before relying on it for client work.
Write down the answers for each approved tool. If the firm is ever asked how it concluded a tool met the statute’s access requirement, that record is the answer.
Configuration supports that conclusion without settling it. An enterprise plan, training turned off, single sign-on, and tight sharing settings are the technical groundwork, but none of them, alone or together, is a guarantee that a tool satisfies the restricted-access requirement. Whether the firm’s setup meets the statute is a judgment for the attorneys and their ethics counsel, made with the technical facts in hand.
Shadow AI Is the Biggest Exposure
One scenario deserves particular attention because it is hard to reconcile with the confidentiality rule: an attorney or paralegal pasting a client document into a personal ChatGPT, Claude, or Gemini account on their own. The firm has no contract with that vendor, no admin controls, no say over retention, and often no idea it’s happening. A firm can’t show that access to those inputs was restricted, because it never controlled them in the first place.
That’s why an inventory of what people actually use has to come first. Shadow AI is already inside most organizations, and staff often turn to personal accounts when firm-approved tools hit usage limits or don’t do what they need. Approving a good tool and giving people enough capacity to use it is part of compliance.
The Firm Setup Checklist Before January 1
- Inventory the AI tools people actually use.Include personal accounts, browser extensions, meeting transcription tools, and AI features built into software the firm already licenses. Ask directly and make it safe to answer honestly.
- Choose approved tools and plans, then configure them.Turn off training on inputs, set retention, restrict workspace sharing, connect single sign-on, and record the answers to the access questions above for each tool.
- Define what can go into each tool.Map the firm’s data types, including the statute’s list of personal identifying information, to each approved tool. Our guide to what client data can go into AI walks through this by data type.
- Build verification into the workflow.Set a review step for AI-assisted work product. In the filing workflow, add a citation check that records which attorney personally verified the citations and when, so the record exists before the document is filed.
- Have counsel decide the disclosure approach, then build the tracking.What disclosure the statute requires is a legal question for the firm’s attorneys and ethics counsel, and the law leaves it unsettled. Once they decide, the technology side is recording AI use on each matter and document so the responsible attorney has what they need to make the disclosure.
- Write the firm AI policy and train everyone.Cover approved tools, permitted inputs, verification, disclosure, and what to do if client information goes into the wrong tool. Train attorneys and staff, and keep a record of who completed it.
- Review after January 1.SB 574 directs the Judicial Council to revisit its standard on generative AI, and State Bar guidance will keep developing. Put a review date on the calendar.
What the Law Leaves Open
Several questions won’t be settled by January 1, and the firm’s policy should note how it’s handling each one in the meantime.
What counts as “delegating” the practice of law. The statute doesn’t define it. Holland & Knight notes that firms with heavily automated workflows will need meaningful attorney review at substantive points.
What triggers court disclosure, and in what form. Drafting an argument with AI and using it to fix grammar may not be treated the same, but the statute doesn’t distinguish them.
Who the “responsible” attorney is on a filing with multiple signers. Both Holland & Knight and Baker McKenzie flag this as unresolved.
What “restricted” access means for a given AI system. Baker McKenzie notes the statute doesn’t define it, so how it applies to different products remains to be determined.
Watch for guidance from the State Bar and the Judicial Council, and confirm the firm’s interpretations with ethics counsel before relying on them.
Technology and governance guidance, not legal advice. Vision Quest Cyber is a technology and AI governance partner, not a law firm. This article summarizes SB 574 based on the chaptered bill text and published law firm analyses, and describes the technical and operational steps firms commonly take. It does not interpret the statute for your firm’s circumstances. Confirm your obligations with qualified ethics counsel. Vendor terms and product features change; verify current settings and contract terms for every tool you approve. Last reviewed: October 2026.
Get your firm’s AI tools and policy ready for January 1
In 15 minutes, we’ll talk through which AI tools your attorneys and staff are using today, what needs to be configured or replaced, and how to put a firm policy in place before the deadline. Learn more about our AI governance services and our work with law firms.
Schedule a 15-Minute CallSources: SB 574 (Umberg), Chapter 858, Statutes of 2026, chaptered text: legiscan.com; bill status: leginfo.legislature.ca.gov • Holland & Knight, “California Enacts Rules Governing Lawyers’ Use of Generative AI” (Oct. 1, 2026): hklaw.com • Baker McKenzie, “California Imposes New AI Rules on Lawyers” (Oct. 13, 2026): bakermckenzie.com
