California SB 574's AI Rules Reach Newsom's Desk With Seven Days Left — and Arbitrators Are In Them

Newsom has seven days to decide on the first US statute that tells arbitrators how to use AI

California SB 574 passed the Legislature on 31 August 2026 without a dissenting vote, and Governor Gavin Newsom has until 30 September 2026 to sign it, veto it, or let it become law unsigned — a deadline that now falls inside a single week. The attention the bill has drawn has been almost entirely about lawyers and hallucinated citations, but the provision with the longest reach is a new Code of Civil Procedure section 1282.1 telling neutral arbitrators that they may not delegate any part of the decisionmaking process to a generative artificial intelligence tool.

That is the part practitioners in this field should read first. Every other state's response to AI in litigation has so far come through standing orders, bar ethics opinions and court rules. If SB 574 is enacted, California becomes the first United States jurisdiction to put duties on arbitrators — and on certified alternative dispute resolution providers — in a statute, enforceable through the machinery that already governs arbitral conduct in this state. The California SB 574 AI rules would apply to a private neutral deciding a consumer claim in Los Angeles exactly as they apply to a lawyer filing a brief in Sacramento.

This piece sets out what the bill actually does, what section 1282.1 would require of arbitrators, the award-level consequence question that the bill does not answer, whether the Federal Arbitration Act (FAA) leaves room for it, and what it means for the provider programs — the American Arbitration Association (AAA) and JAMS among them — that are already building AI into arbitration administration.

What does California SB 574 do?

SB 574, authored by Senator Tom Umberg with Assemblymember Macedo as coauthor, imposes a statutory framework on generative AI use across five points in the California legal system: practising attorneys, court filings, judicial administration, arbitrators, and certified ADR providers. It does this through five code changes.

Code section	Action	What it addresses
Bus. & Prof. Code § 6068.1	Added	Attorney duties: no delegation of the practice of law to generative AI; no entry of confidential, personal identifying or other nonpublic information into a generative AI system unless access is restricted to the attorney and persons obligated to protect it; reasonable steps to verify accuracy, including all case and statutory citations; correction of erroneous or hallucinated output
Bus. & Prof. Code § 6173	Amended	Certified alternative dispute resolution firms, providers and practitioners under the State Bar's certification program
Code Civ. Proc. § 128.7	Amended	A filed paper may not contain a citation that has not been personally read and verified by an attorney responsible for the filing, including a citation supplied by generative AI
Code Civ. Proc. § 180	Added	Directs the Judicial Council to publicly revisit, and revise as necessary, Standard 10.80 of the California Standards of Judicial Administration to reflect the further development of generative AI
Code Civ. Proc. § 1282.1	Added	Arbitrator duties: no delegation of any part of the decisionmaking process to generative AI; no substitution for independent analysis; disclosure before relying on AI-generated material outside the record; full responsibility for all aspects of an award

The attorney provisions have drawn the commentary because they are the ones that bite on daily practice. The section 128.7 amendment is the sharpest of them: as several analyses have read the text, the certification is violated where the signing attorney did not personally read and verify a cited source, regardless of whether the citation turns out to be accurate and regardless of whether AI produced it. That converts a diligence expectation into a per-filing certification, and it is the reason the loudest objection to the bill is a cost objection rather than a policy one.

Section 180 is a single directive with no prescribed substance and no deadline: the Judicial Council is told to revisit Standard 10.80, not what to conclude. Section 6173 touches the State Bar's existing certification program for alternative resolution firms, providers and practitioners, which is how the bill reaches ADR institutions rather than only individual neutrals.

What does SB 574 require of arbitrators who use generative AI?

Section 1282.1 would impose four duties on a neutral arbitrator, and they are cumulative rather than alternative.

No delegation of decisionmaking. An arbitrator shall not delegate any part of the decisionmaking process to any generative AI tool. The word "any" is doing real work: this is not a prohibition on outcome-generation alone, and on its face it reaches intermediate steps — issue framing, evidentiary weighing, drafting of reasoning — not merely the disposition.

No substitution for independent analysis. Use of generative AI tools shall not replace the arbitrator's independent analysis of the facts, the law and the evidence. This is the standard against which any AI-assisted step would be measured if challenged, and it is framed as a floor the arbitrator must meet personally.

Disclosure before relying on out-of-record AI material. An arbitrator is barred from relying on AI-generated information outside the record without first disclosing that use to the parties and, as far as practical, giving them an opportunity to comment. The arbitrator must not assume that sources exist or are fairly described, and the sources must be verifiable independent of the AI tool. This is a genuine Due Process (Arbitral) provision dressed as a technology rule — it is the familiar prohibition on deciding by reference to material the parties never saw, applied to a new source of such material.

Full responsibility for the award. The arbitrator must assume full responsibility for all aspects of an Arbitral Award regardless of any AI use. A savings clause preserves the arbitrator's existing duty of neutrality and duty to exercise reasonable competence and diligence, so the section adds to rather than displaces those obligations.

What section 1282.1 does not do is tell anyone what happens next when it is broken. That silence is the most consequential feature of the provision, and no published analysis has worked it through.

Can an arbitration award be vacated if an arbitrator violates SB 574?

The bill does not say, and a party who wants relief would have to fit the breach into the grounds that already exist. Under Code of Civil Procedure section 1286.2(a), a California court shall vacate an award where it was procured by corruption, fraud or other undue means; where there was corruption in any of the arbitrators; where the party's rights were substantially prejudiced by misconduct of a neutral arbitrator; where the arbitrators exceeded their powers and the award cannot be corrected without affecting the merits; or where the party's rights were substantially prejudiced by a refusal to postpone on sufficient cause, a refusal to hear material evidence, "or by other conduct of the arbitrators contrary to the provisions of this title."

That last clause is the obvious hook. Section 1282.1 would sit inside Title 9 of Part 3 of the Code of Civil Procedure — the title that clause refers to — so an arbitrator who decided part of a case with a generative AI tool, or who relied on AI-generated authority the parties never saw, would be engaged in conduct contrary to a provision of that title. On the statutory language, that is a cleaner route to Vacatur than the corruption grounds, which carry a scienter freight that an AI-assistance dispute will rarely bear. It is not automatic: the clause requires substantial prejudice, which means a challenger must show the AI use affected the outcome and not merely the workflow.

Two consequences follow for anyone drafting or defending awards in California. The first is evidentiary: substantial prejudice is close to unprovable without some record of how the award was produced, which pushes the fight toward discovery into the arbitrator's process — territory California courts have historically kept closed. The second is behavioural: a neutral who wants to foreclose the argument entirely will simply not use the tools, which is very likely the bill's practical effect regardless of what its text technically permits.

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

Probably in substantial part, but the question is open and nobody has litigated it. The doctrinal line is familiar: section 2 of the Federal Arbitration Act (FAA) withdrew the power of states to require a judicial forum for claims the parties agreed to arbitrate, and FAA Preemption reaches state rules that single out arbitration agreements or obstruct their enforcement. State rules that govern how an arbitration is conducted, without affecting whether the agreement is enforced, sit outside that preemptive core.

Section 1282.1 is on the conduct side of that line. It does not condition enforcement of an Arbitration Clause, does not add a precondition to compelling arbitration, and does not make any category of claim non-arbitrable. It regulates what a neutral may do once seated. That framing gives it a materially better preemption posture than California's more familiar arbitration statutes — the fee-payment provisions at Code of Civil Procedure sections 1281.97–1281.98, for instance, which attach a forfeiture consequence to conduct and have drawn preemption challenges precisely because of it.

The pressure point is the remedy rather than the rule. For an award governed by the FAA, the federal grounds for vacating an award are narrow and are set out in FAA section 10; a state statute that in practice supplies an additional route to set aside an FAA-governed award invites the argument that it is doing what the FAA does not permit. A respondent facing a section 1282.1 challenge to an FAA-governed award will make that argument. Until a court rules, the safest reading for practitioners is that the conduct duties apply to arbitrations seated in California and the award consequences are contested where the FAA governs.

What SB 574 means for AAA, JAMS and institutional AI adjudication

The timing is the story. The AAA-ICDR announced on 17 September 2025 that it would release an AI Arbitrator, which became available from November 2025 for two-party, documents-only construction cases. It is opt-in, was trained on more than 1,500 construction awards and calibrated with human arbitrator input, and is structured with human oversight at two levels: the parties review and validate the AI's understanding of their submissions, and a human arbitrator then oversees, revises where needed, confirms and authorises each outcome before the award issues. The AAA-ICDR has said the tool will extend to further industries, case types and higher-value claims.

That architecture was designed for exactly the objection section 1282.1 codifies. Because a human neutral confirms and authorises every award, an AAA-ICDR AI Arbitrator proceeding is not obviously a delegation of "any part of the decisionmaking process" — but nor is it obviously outside the phrase, because the drafting of reasoning and the weighing that precedes confirmation are parts of that process on any natural reading. Whether the programme could be offered in a California-seated case after 1 January 2027 without an express party agreement is a live question the bill's text does not settle, and the opt-in structure may prove to be the answer: parties who consent in writing to an AI-native process are in a different position from parties whose neutral quietly uses a tool.

The institutional guidance landscape around this is already dense. The AAA-ICDR issued guidance on arbitrators' use of AI tools in March 2025, built on four principles: accuracy and reliability of AI-generated information, fairness and due process, preservation of independent decisionmaking, and transparency with the parties. The Silicon Valley Arbitration & Mediation Center published guidelines on the use of AI in arbitration on 30 April 2024. JAMS published its Artificial Intelligence Disputes Clause and Rules effective 14 June 2024 — those govern disputes about AI systems rather than neutrals' use of AI, and they apply only where the parties adopt them, which is a distinction worth keeping straight when the two are cited together.

SB 574 would convert the overlapping middle of that guidance into California law, enforceable by a court rather than by an institution. Incorporation of Provider Rules would no longer be the whole of the answer to what a neutral may do.

Where mass arbitration workflows sit under section 1282.1

This is where the provision reaches furthest and where no competing analysis has looked at all. Mass Arbitration administration is the single most obvious candidate for AI assistance anywhere in dispute resolution: thousands of near-identical demands, standardised eligibility questions, and Claimant Vetting exercises that turn on document checks rather than judgment.

Under the AAA's Mass Arbitration Supplementary Rules, which apply to mass arbitrations commenced after 15 January 2024, a Process Arbitrator is appointed to resolve administrative disputes across the batch — filing-requirement compliance, fee allocation, whether demands meet the rules. The fee structure replaced per-case filing fees with a single Initiation Fee, set at $3,125 for claimants and $8,125 for a business respondent. Respondents have been rewriting clauses around this machinery: Amazon's reinstated US Conditions of Use, effective 14 August 2026, treats 25 or more demands raising the same issue within six months as a mass arbitration and directs them into batches of at least 25.

Section 1282.1 speaks to "an arbitrator." A process arbitrator is an arbitrator. On the face of the statute, a process arbitrator who used a generative AI tool to triage 4,000 demands for compliance with an Evidentiary Precondition would be delegating part of a decisionmaking process, and the fact that each individual decision is administrative rather than merits-based is not a distinction the text draws. Providers running California-seated mass filings after 1 January 2027 would need either a defensible account of why batch administration is not "decisionmaking," or express party agreement, or human sign-off on every determination — which is precisely the cost structure that mass arbitration procedures were built to avoid.

Does SB 574 require lawyers to disclose that they used AI in a court filing?

The coverage conflicts on this, and the conflict matters. Several news write-ups state that the bill would require lawyers to disclose when they use AI to prepare documents submitted in court. The law-firm analyses of the operative text describe the attorney duties differently — as non-delegation, confidentiality, verification of accuracy including all case and statutory citations, and correction of erroneous or hallucinated output — with the section 128.7 amendment framed as a personal-verification certification rather than an affirmative disclosure requirement. Separate commentary describes disclosure of AI-generated content to the public as a matter for firm policy rather than statutory command.

Where the bill is unambiguous about disclosure is in section 1282.1, and the duty there runs from the arbitrator to the parties, not from counsel to the court: no reliance on out-of-record AI-generated information without prior disclosure and, as far as practical, an opportunity to comment. Practitioners building compliance policies in the next 92 days should assume the verification and correction duties are certain, treat an attorney court-disclosure obligation as unsettled until the enrolled text is read against the final amendments, and treat the arbitrator disclosure duty as settled.

Compliance timeline if the bill is signed

Date	Event
31 August 2026	Legislature passes SB 574 without a recorded no vote; Assembly floor tally reported at 75-0
23 September 2026	Bill unsigned; seven days remain on the governor's action window
30 September 2026	Deadline to sign or veto; absent action, the bill becomes law without signature
1 January 2027	Operative date, 92 days after the deadline
No fixed date	Judicial Council revisits Standard 10.80 under new Code Civ. Proc. § 180

What it means for litigators, respondents' counsel and claimants' firms

For California litigators. The section 128.7 exposure is per filing and personal. Firms that route research through AI tools need a verification step that produces a record of who read which source, because the certification is the attorney's and the sanction runs against the attorney. The State Bar disciplinary track sits alongside section 128.7 sanctions, so the same conduct can be addressed twice.

For respondents and clause drafters. Every California-seated arbitration clause drafted after enactment invites a choice the drafter did not previously have to make: say nothing about AI and inherit section 1282.1's default, or address the neutral's use of AI expressly, including consent to an institutional AI-assisted process. Clauses that incorporate provider rules by reference now incorporate whatever those providers do with AI, subject to a statutory floor.

For claimants' and mass filing counsel. The practical effect cuts both ways. Section 1282.1 gives a claimant a new argument against an adverse award produced with undisclosed AI assistance, which is a real lever in a jurisdiction where award challenges usually fail. It also raises the administrative cost of the very batch processes that make mass filings viable, and cost that lands on the provider lands eventually on the parties.

For personal injury and consumer firms outside California. California's rules travel. Legislatures in other states have tended to follow its lead on legal-practice regulation, and the professional-responsibility content here — verify citations, do not delegate judgment, do not feed client confidences into a public model — is portable regardless of what Newsom does by 30 September.

FAQ

Has Newsom signed SB 574?

Not as of 23 September 2026. The governor's deadline to sign or veto is 30 September 2026, and if he takes no action the bill becomes law without his signature.

When would SB 574 take effect?

If enacted, SB 574 would take effect on 1 January 2027, following the ordinary California rule for bills passed in the regular session.

What are the penalties for violating SB 574?

Attorney violations are addressed through sanctions under Code of Civil Procedure section 128.7 and through the State Bar disciplinary process. The bill does not specify a penalty for an arbitrator who breaches section 1282.1, which is why the vacatur analysis above matters.

Does SB 574 ban lawyers from using ChatGPT or similar tools?

No. It prohibits delegating the practice of law to generative AI, prohibits entering confidential, personal identifying and other nonpublic information into a system without restricted access, and requires verification of outputs and correction of hallucinated material. Assisted work that the attorney independently verifies is not prohibited.

Would SB 574 stop the AAA-ICDR AI Arbitrator being used in California?

The bill does not name any programme. Whether an opt-in, human-confirmed AI-assisted process constitutes delegation of "any part of the decisionmaking process" under section 1282.1 is unresolved on the text, and express party agreement is the most likely route by which providers would seek to place such proceedings beyond argument.