California SB 574 Sets Binding AI Rules for Lawyers

Estimated reading time: 7 minutes · Last updated: 12 November 2026

California's Senate Bill 574, signed by Gov. Gavin Newsom on 30 September, writes limits on generative AI into state law for attorneys and arbitrators. Lawyers may not hand the practice of law to an AI system. They must check every case and statutory citation it produces and tell courts when they used it. Arbitrators may not let AI make any part of a decision. In a 26 October analysis, attorneys at Mayer Brown call SB 574 the first US statute aimed specifically at these legal uses of AI. The rules start on 1 January. Two practical questions remain open: what counts as delegation, and whether duties written for courts also reach arbitration.

What SB 574 Requires of Lawyers Before They File

The core rule for lawyers is easy to state. Generative AI can help, but it cannot practise law. Generative AI is software that produces new text, such as a draft brief, in response to a prompt. Under SB 574, an attorney may not hand the practice of law over to such a tool.

Verification comes next. Attorneys must take reasonable steps to check what the system produces, each case and statutory citation included, and must fix anything that is wrong or invented. Mayer Brown links this duty to a familiar failure: AI tools citing authorities that sound real and do not exist.

The filing rules go further. A document submitted to a court may cite only authorities that the responsible attorney has personally read and verified. A citation that merely looks credible is not enough. Lawyers must also disclose generative AI use in what they file with courts. Mayer Brown notes that many of these duties already exist as ethical obligations, and the statute now gives them a firmer legal footing.

Why Client Data Now Needs a Locked Door

SB 574 also governs what goes into an AI system in the first place. Attorneys may not feed a tool confidential material, personally identifying details or other nonpublic information unless access is confined to the attorney and authorised people who are bound to protect it. A tool whose access cannot be limited in that way does not meet the standard.

This provision turns a vendor's access controls into a legal question. It is the part of the law most likely to land on a firm's technology decisions rather than its brief-writing habits, because it concerns where data sits, not how a lawyer reasons.

Arbitration adds further layers. Confidentiality there can arise from the parties' own agreement, from an institution's rules or from a procedural order, so the body of protected material may be wider than in court. Mayer Brown's attorneys treat a review of each tool's access controls as an essential step in protecting case files.

How the Law Keeps Arbitrators' Decisions Human

For arbitrators, the focus is the decision itself. A new section of the California Arbitration Act bars handing any part of the decision-making process to generative AI. The ban covers the whole process, not only the final ruling.

A second rule deals with information from outside the record. An arbitrator who wants to rely on AI-generated material the parties never put in evidence must first disclose it and, where practical, let the parties comment. The aim is procedural fairness, so that a party can challenge anything that shapes an award. Existing duties of neutrality, competence and diligence still apply.

Sen. Tom Umberg, the bill's author, said decisions entrusted to arbitrators should be "made by people, not algorithms," in remarks reported by Law.com on 1 October. The principle has a parallel abroad. In ARIHQ v. Santé Québec, the Quebec Superior Court set aside an arbitral award because decision-making authority had been improperly delegated to AI, Mayer Brown notes.

Where Machine-Drafted Awards Meet the New Rules

California's line runs into a service that already exists. In November 2025, the American Arbitration Association's International Centre for Dispute Resolution launched an AI Arbitrator built to generate awards, not just to support a human tribunal. It is open only to parties who consent, and only in narrowly defined construction disputes decided on documents alone. The system drafts an award that people then review, validate and revise.

Whether that human layer satisfies a rule demanding human decision-making is still unsettled, Mayer Brown's attorneys write. A draft produced by software and approved by a person sits squarely in the gap the statute leaves undefined.

Two further gaps matter. The statute does not spell out what counts as delegating legal practice, and it is unclear whether the court-focused disclosure and filing rules also cover submissions in arbitration. Mayer Brown says the answers could decide where firms draw the line between permitted assistance and prohibited substitution.

What Lawyers and Arbitrators Should Do Before 1 January

Mayer Brown recommends that attorneys and arbitrators audit their AI practices before 1 January, the date SB 574 takes effect. That means recording how each service is used and confirming that confidential material stays inside restricted systems.

Parties to disputes are advised to agree AI protocols at the start of a case and raise expectations with the tribunal early, rather than discovering a disagreement once an award has been issued.

The consequences reach beyond an embarrassing brief. Mayer Brown warns that an award revealing undisclosed AI involvement, or improper delegation, could be challenged on statutory grounds as well as on procedural fairness. For a firm, the practical shift is from trusting individual judgment to being able to show, on paper, where AI touched the work and where a named person took responsibility for it.

Will SB 574 hold up in practice?

The bottom line

SB 574 does less to invent new duties than to harden old ones. Verifying citations, guarding client data and keeping judgment human were already expected of California lawyers. From 1 January those expectations are written into statute, and arbitrators are covered too. The harder work lies at the unsettled edges. The law does not define where assistance ends and delegation begins, and the AAA's AI Arbitrator tests that boundary directly with machine-drafted awards. The Quebec Superior Court has already shown that an award can be set aside over improper delegation, and California has now given parties a statutory basis for the same argument. Firms that can document how AI entered a file will be better placed than those that cannot.

What to watch

• SB 574 takes effect on the coming 1 January, the first point at which lawyers and arbitrators in California can be held to its rules.
• Watch for any guidance or ruling on whether the court-focused disclosure and filing duties also cover arbitral submissions; no date has been set.
• Watch whether the AAA's International Centre for Dispute Resolution adjusts its AI Arbitrator for California disputes; no announcement has been made.