Newsom Signs SB 574: California's AI Rules for Lawyers and Arbitrators Take Effect January 1, 2027

Newsom signs SB 574: the first state law on AI use by lawyers and arbitrators

Governor Gavin Newsom signed SB 574 on Wednesday, September 30, 2026, which makes California the first state to set statutory rules for how licensed attorneys and arbitrators use generative artificial intelligence. Starting January 1, 2027, a California lawyer must disclose generative AI use to the court for documents submitted to it and must personally verify every citation in a filing. An arbitrator may not delegate any part of the decisionmaking process to an AI tool.

Newsom signed SB 574 on the last day he had to act. It was one of 13 AI bills he signed that day, most of them aimed at the workplace. The other bills require human review before AI-driven firing or discipline, require notice when AI causes mass layoffs, and limit workplace surveillance and deepfakes. SB 574 is the one aimed at the legal profession. Senator Thomas Umberg (D-Santa Ana) titled it the Court A.I. Protection Act. It passed without a recorded no vote: the Assembly vote on August 31, 2026 was 75-0. The Senate Judiciary Committee analysis listed the bill as author-sponsored, with Oakland Privacy in support and no known opposition.

Most coverage of the signing has been about lawyers and hallucinated citations. The signature also turns two open questions into settled law for this field. Court disclosure of AI use is mandatory, not optional. And Code of Civil Procedure section 1282.1, the arbitrator provision, is now a statute that California-seated neutrals, provider programs and mass-arbitration administrators have three months to plan around. This piece covers each operative section, who it binds, and what still needs to be litigated.

Did Newsom sign SB 574, and when does it take effect?

Yes. Newsom signed SB 574 on September 30, 2026, and it takes effect on January 1, 2027, the default operative date for a regular-session California bill with no urgency clause. That leaves 93 days between signature and the date the duties apply.

The legislative history explains why the final text holds up. Umberg introduced SB 574 on February 20, 2025. It was amended twice in the Senate and four times in the Assembly, cleared both houses on August 31, 2026, the final day of the session, and was enrolled and presented to the governor on September 9, 2026. Each round of amendment narrowed and clarified the drafting. Several commentators on the earlier versions disagreed about how far the disclosure duty went. The enacted text answers that question, as discussed below.

Signature does not change any rule that was already binding. California Rules of Court, rule 10.430, adopted by the Judicial Council in 2025, required every court that allows staff or judicial officers to use generative AI to adopt a use policy by December 15, 2025. Standard 10.80 sets guidelines for judicial officers' use of AI within their adjudicative role. SB 574 adds new Code of Civil Procedure section 180, which directs the Judicial Council to revisit and, as needed, revise Standard 10.80 as the technology develops. It sets no deadline and prescribes no content. The new statutory duties fall on the bar and on arbitrators, not on judges.

What does SB 574 require lawyers to do when using AI?

SB 574 requires California lawyers to keep the legal judgment in their own hands, keep client confidences out of unrestricted AI systems, verify and correct everything an AI tool produces, avoid discriminatory AI-assisted conduct, and disclose AI use to the court for documents submitted to it. The duties are set out in five code changes:

Code section	Change	What it requires
Bus. & Prof. Code § 6068.1	Added	Attorney may not delegate the practice of law to generative AI; no entry of confidential, personal identifying or other nonpublic information into a system unless access is limited to the attorney and persons authorized under confidentiality obligations; reasonable steps to verify accuracy and correct erroneous or hallucinated output; no use that unlawfully discriminates against or disparately impacts protected classes; disclosure of generative AI use to the court for all documents submitted to the court; consider disclosure for public-facing content
Code Civ. Proc. § 128.7	Amended	No filed paper may contain a citation that the responsible attorney has not personally read and verified, whether it came from AI or any other source
Code Civ. Proc. § 1282.1	Added	Arbitrator may not delegate any part of the decisionmaking process to generative AI; AI may not replace independent analysis; disclosure before relying on out-of-record AI material; arbitrator bears full responsibility for the award
Bus. & Prof. Code § 6173	Amended	Extends the framework to alternative dispute resolution providers and practitioners certified by the State Bar
Code Civ. Proc. § 180	Added	Judicial Council to revisit Standard 10.80 on judicial officers' use of generative AI

The non-delegation rule in section 6068.1 is the core provision, and it does not ban the tools. As Holland & Knight's analysis of the enacted law notes, the statute does not list tasks that may or may not be done with AI help. The natural reading is that research, first-draft writing, document review, discovery triage and correspondence can all be AI-assisted, provided the lawyer supplies the legal judgment and stands behind the result. The statute prohibits handing an AI agent the case and accepting its output as the work product.

The confidentiality duty is the one most likely to change firm technology budgets. The test is access, not vendor reputation. An enterprise deployment where inputs are restricted to the firm and its confidentiality-bound agents qualifies. A consumer chatbot account where the provider may retain or train on inputs does not, unless its terms say otherwise. The duty covers "personal identifying" and "other nonpublic" information as well as privileged material. For a personal injury practice, that covers medical records, Social Security numbers, accident reports and settlement figures, which make up most of its files.

The anti-discrimination duty gets less attention but is in the text. Attorneys must ensure their AI use does not unlawfully discriminate against or disparately impact people based on characteristics protected by federal or state law. This matters most where AI screens clients, scores case value or drafts intake decisions. Mass tort and mass-arbitration intake operations increasingly run on automated questionnaires, so this provision applies directly to them.

Do California lawyers have to disclose AI use to the court?

Yes. Under the enacted section 6068.1, an attorney who uses generative AI must disclose that use to the court for all documents submitted to the court. For content provided to the public, such as marketing, client alerts and website copy, the attorney must consider whether to disclose, but disclosure is not mandatory.

This is the change from the pre-signature coverage. While the bill sat on Newsom's desk, some analyses said it only required lawyers to consider disclosure, others said disclosure to the court was mandatory, and our own September 23 piece called the court-disclosure question unsettled. Post-enactment analysis of the statute, including Holland & Knight's, reads it as a mandatory court-disclosure rule paired with a discretionary public-disclosure rule. The Senate Judiciary Committee's analysis described it the same way.

The statute does not prescribe a form of disclosure. Expect three things before January 1: firms will adopt standard certification language, courts will issue local rules or standing orders specifying form and placement, and the first fights will be over what counts as "use." A brief drafted from scratch by a lawyer who used an AI research tool to find leads differs from a brief whose first draft an AI wrote. On its face, the statute's disclosure duty applies to both, because it turns on use of generative AI for the document and not on how much of the text the AI wrote. Until courts say otherwise, the safer practice is broad disclosure in a short, standard sentence.

The citation rule: section 128.7 and the Noland sanction

The section 128.7 amendment applies to every filing, AI or not. A paper filed in a California court may not contain a citation that the attorney responsible for the filing has not personally read and verified. As Clearly Concise's compliance guide notes, the duty covers citations from an associate, a paralegal, a research service, a prior brief or any other source, not only AI.

The wording closely follows California's leading sanctions opinion. In Noland v. Land of the Free, L.P. (No. B331918), decided September 12, 2025, Division Three of the Second District Court of Appeal found that 21 of the 23 case quotations in an appellant's opening brief were fabricated by generative AI tools. Counsel had run the brief through ChatGPT, Claude, Gemini and Grok, hoping each would catch the others' errors. The court imposed a $10,000 sanction, referred counsel to the State Bar, and wrote that no filed paper "should contain any citations—whether provided by generative AI or any other source—that the attorney responsible for submitting the pleading has not personally read and verified." SB 574 turns that sentence from a published admonition into a certification duty under section 128.7.

What changes in practice is who must do the reading. Before SB 574, a supervising partner who signed a brief built on an associate's citations was exposed to sanctions only if the citations were wrong. After January 1, 2027, the text supports reading the signature as a certification that a responsible attorney personally read each cited source. That stands on its own: the certification can be violated even if every citation turns out to be accurate. Farella Braun + Martel partner Jeff Fisher has said this is the one part of the bill that goes too far, because of what it adds to the cost of litigation.

Can arbitrators use AI under California law?

Yes, but only as an assistant. From January 1, 2027, section 1282.1 bars an arbitrator from delegating any part of the decisionmaking process to a generative AI tool, and from relying on AI-generated information outside the record without first disclosing that use to the parties and, as far as practical, letting them comment. The section imposes four cumulative duties on the neutral:

• No delegation. No part of the decisionmaking process may be delegated to generative AI. "Any part" covers intermediate steps such as framing issues, weighing evidence and drafting the reasoning, not just the final ruling.
• Independent analysis. AI use may not replace the arbitrator's own analysis of the facts, the law and the evidence.
• Disclosure before reliance. AI-generated material from outside the record may not be relied on until the parties have been told and, where practical, heard. This is a Due Process (Arbitral) rule: it prohibits deciding a case on material the parties never saw, applied to a new source of such material.
• Full responsibility. The arbitrator is responsible for all aspects of the Arbitral Award regardless of AI use. Existing duties of neutrality, competence and diligence are expressly preserved.

Nothing in section 1282.1 stops a neutral from using AI to organize exhibits, build a chronology, check a draft for internal consistency or search the record. What it rules out is letting the tool decide, or bringing in AI-generated information the parties have not seen. The amended section 6173 reaches the provider level through the State Bar's certification program for ADR firms and practitioners, so institutions as well as individual arbitrators now have a statutory standard.

Can an award be vacated if the arbitrator broke the AI rule?

SB 574 does not say. The likeliest route is the existing vacatur statute. Code of Civil Procedure section 1286.2(a)(6) requires a court to vacate an award where a party's rights were substantially prejudiced by, among other things, "other conduct of the arbitrators contrary to the provisions of this title." Section 1282.1 sits in that same title. An arbitrator who let an AI tool write the reasoning, or who relied on undisclosed AI-generated authority, would be acting contrary to a provision of the title.

Two limits make this narrower than it sounds. First, the ground requires substantial prejudice, so a challenger must show the AI use affected the result, not just the workflow. Second, proving how an award was produced usually requires discovery into the neutral's process, which California courts have long been reluctant to allow. The likeliest early cases will involve an obvious trace: a fabricated citation in a reasoned award, or reasoning that relies on material outside the record. Vacatur stays an uphill motion, but section 1282.1 gives losing parties a concrete statutory hook they did not have before.

Does SB 574 apply to arbitrations under the Federal Arbitration Act?

The conduct rules most likely apply to California-seated arbitrations even where the Federal Arbitration Act (FAA) governs the agreement, but the remedy is contestable. FAA Preemption reaches state rules that refuse to enforce arbitration agreements, single them out for disfavor, or interfere with fundamental attributes of arbitration. Section 1282.1 does none of those things directly. It does not condition enforcement of an Arbitration Clause, make any claim non-arbitrable, or add a precondition to compelling arbitration. It regulates what a neutral may do once appointed, and that kind of conduct rule is generally on the safer side of the preemption line.

The pressure point is vacatur. For an award governed by the FAA, the federal grounds in FAA section 10 are narrow and, under Hall Street Associates v. Mattel, exclusive for federal-court review under the FAA. A respondent defending an FAA-governed award against a section 1282.1 challenge will argue that a state statute cannot add a ground for setting it aside. A claimant will answer that undisclosed AI decisionmaking fits within existing FAA grounds such as misconduct or exceeding powers. No court has ruled. Until one does, practitioners should assume that the conduct duties bind neutrals sitting in California and that the remedy is contested wherever the FAA governs.

What SB 574 means for AAA, JAMS and mass arbitration

The provider question is now operational. The American Arbitration Association (AAA), through the AAA-ICDR, announced its AI Arbitrator in September 2025 for two-party, documents-only construction cases. Use is opt-in, and a human arbitrator reviews, revises where needed and confirms every outcome before an award issues. That human confirmation is the program's best argument under section 1282.1. Its weak point is the phrase "any part of the decisionmaking process," because drafting and preliminary weighing are parts of that process on any natural reading. Express written party consent to an AI-assisted process is the most likely way providers will offer such products in California-seated cases after January 1. Whether consent overrides a statutory duty, rather than defining its scope, is a question the text leaves open.

The larger exposure is in Mass Arbitration administration. Under the AAA's Mass Arbitration Supplementary Rules, a Process Arbitrator decides administrative issues across a batch: whether demands meet filing requirements, how fees are allocated, and whether claimants have satisfied pre-filing conditions. JAMS's Mass Arbitration Procedures and Guidelines, effective May 1, 2024, also send threshold and procedural questions through centralized steps before individual merits hearings. These are the high-volume, document-checking decisions where AI triage saves the most money. Section 1282.1 says "an arbitrator," and a process arbitrator is one. The statute does not distinguish administrative decisions from merits decisions. A process arbitrator who used an AI tool to sort 5,000 demands into compliant and deficient piles, and then signed the result, is exposed to the argument that part of the decisionmaking was delegated.

Three ways to comply are open: human sign-off on each determination, which removes much of the efficiency; express consent to AI-assisted triage in the arbitration agreement or a procedural order; or a reading that limits section 1282.1 to merits adjudication, which the text does not support. Respondents redrafting consumer and employment clauses under the Incorporation of Provider Rules model should note that the clause now incorporates a statutory floor as well as the provider's rules.

What happens if a lawyer violates SB 574?

A lawyer who violates SB 574 faces court sanctions under Code of Civil Procedure section 128.7 for filing violations and State Bar discipline for breaches of section 6068.1. The same conduct can trigger both. The law gives clients no private right of action. As DK Law, a California personal injury firm, noted in its client explainer, SB 574 tells attorneys what they may not hand to a machine but gives clients no direct way to enforce it. Clients can still file State Bar complaints and bring malpractice claims, where breach of a statutory duty is likely to be cited as evidence of breach of the standard of care.

Arbitrators face a less complete enforcement scheme. There is no sanction provision in section 1282.1 itself. The consequences come through the vacatur analysis above, through the State Bar's ADR certification program for certified providers, and through provider-level removal and disclosure processes. Practitioners should expect challenge motions and arbitrator-disclosure requests to start asking about AI use as a matter of course.

Compliance timeline: from signature to the first challenges

Date	Event
February 20, 2025	SB 574 introduced by Senator Umberg
September 12, 2025	Noland v. Land of the Free: $10,000 sanction for AI-fabricated quotations; language later tracked in § 128.7
December 15, 2025	Deadline under rule 10.430 for courts permitting generative AI to adopt use policies
August 31, 2026	Legislature passes SB 574; Assembly vote 75-0; no recorded no votes
September 9, 2026	Bill enrolled and presented to the governor
September 30, 2026	Newsom signs SB 574 with 12 other AI bills
January 1, 2027	§§ 6068.1, 128.7 (as amended), 1282.1, 6173 (as amended) and 180 become operative
No fixed date	Judicial Council revisits Standard 10.80 under § 180

What it means for PI firms, mass-claims counsel, respondents and neutrals

For personal injury and mass tort firms. The cost question is about economics, not ethics. Under a Contingency Fee, every hour spent personally checking citations comes out of the firm's margin, not the client's bill. High-volume practices that file template motions, discovery briefs and Plaintiff Fact Sheet (PFS) disputes should build a verification log now: who read which source, and when. AI vendors handling medical records must pass the access-restriction test in section 6068.1, so procurement contracts should be checked against it before January 1.

For claimants' mass-arbitration counsel. Section 1282.1 cuts both ways. It gives claimants a new challenge to an adverse award produced with undisclosed AI help, and a basis to ask process arbitrators how batch determinations were made. It also threatens the AI-assisted Claimant Vetting and administration that keep high-volume filings cheap. Higher administrative costs tend to raise the settlement value of a well-built filing.

For respondents and clause drafters. Any California-seated clause drafted from now on should address AI use by the neutral expressly, whether by consenting to an AI-assisted process, limiting it, or requiring disclosure. Staying silent means accepting the section 1282.1 default and leaving the FAA preemption fight for the vacatur stage.

For arbitrators and providers. Adopt a written AI-use policy that tracks the four section 1282.1 duties, disclose AI tooling at appointment, and keep a record of how each award was drafted. A neutral who cannot show that independent analysis happened will have a hard time defending an award in 2027.

For lawyers outside California. California is the first state to make these duties statutory, and its legal-ethics rules tend to spread to other states. The duties themselves are already good practice everywhere: verify citations, do not delegate judgment, and keep client confidences out of open AI models.

FAQ

Did Newsom sign SB 574?

Yes. Governor Newsom signed SB 574 on September 30, 2026, as one of 13 AI-related bills he signed that day.

When does SB 574 take effect?

SB 574 takes effect on January 1, 2027. Its duties for attorneys, arbitrators and certified ADR providers apply from that date.

Does SB 574 ban lawyers from using ChatGPT or other AI tools?

No. SB 574 permits AI as an assistant. It prohibits delegating the practice of law to AI, putting confidential or nonpublic information into AI systems without restricted access, and filing unverified output, and it requires disclosure of AI use to the court.

Do lawyers have to tell the court they used AI under SB 574?

Yes. The enacted section 6068.1 requires attorneys to disclose generative AI use to the court for all documents submitted to the court. Disclosure for public-facing content is something the attorney must consider but is not required to make.

Can a California arbitrator use AI to decide a case?

No. Code of Civil Procedure section 1282.1 bars an arbitrator from delegating any part of the decisionmaking process to generative AI and requires disclosure to the parties before relying on any AI-generated material outside the record.

Does SB 574 apply to arbitrations governed by the Federal Arbitration Act?

The arbitrator-conduct rules most likely apply to California-seated arbitrations under the FAA, because they regulate conduct rather than enforceability. Whether a breach can be grounds to vacate an FAA-governed award has not been decided and is likely to be contested.