California Gov. Gavin Newsom signed Senate Bill 574 on Sept. 30, transforming artificial intelligence guidelines for legal professionals and arbitrators into binding statutory rules as fabricated case citations intensify industry-wide scrutiny.
The new statute, which takes effect on Jan., changes the following: 1, represents the first state-level legislation in the United States aimed specifically at regulating generative AI use in legal work, moving past advisory frameworks into mandatory compliance.
Attorneys Must Verify AI Citations and Disclosures
- Attorneys face strict prohibitions against delegating the practice of law to generative AI, alongside mandatory verification protocols for every case and statutory citation.
- Court filings may only contain citations that the responsible attorney has personally read and verified, accompanied by formal disclosures of AI utilization.
- Arbitrators are strictly barred from delegating any part of the decision-making process to automated systems under a new section of the California Arbitration Act.
The legislation addresses two distinct risks identified in legal proceedings: unreliable output entering court records and critical decisions being delegated to software lacking professional accountability. Many provisions reinforce existing ethical duties, but encoding them into state statute creates an explicit legal foundation for enforcement.
Attorneys must take reasonable steps to verify every output generated by automated tools. That requirement directly targets the recurring problem of systems fabricating plausible legal authorities that do not exist.
Furthermore, human accountability extends past confirming whether a generated answer merely appears credible. Documents submitted to court may contain exclusively citations the responsible attorney has personally reviewed, according to an analysis published on Oct. 26 by attorneys at Mayer Brown.
Attorneys must also disclose generative AI use in court submissions. They are prohibited from entering confidential, personally identifying, or nonpublic information into systems unless access remains strictly restricted to authorized individuals bound by confidentiality obligations.
For arbitration practices, confidentiality issues frequently arise from party agreements, institutional rules, or procedural orders. Reviewing a software tool’s access controls has thus become essential for safeguarding case material.
Restrictions placed on arbitrators focus directly on decision-making processes. A new section of the California Arbitration Act prohibits delegating any part of that responsibility to generative AI.
Arbitrators are equally barred from relying on AI-generated information outside the formal record without first disclosing it to the parties and allowing adequate time for comment.
As the measure’s author put it, decisions entrusted to arbitrators should be “made by people, not algorithms,” Sen. Tom Umberg told Law.com on Oct. 1. That safeguard protects procedural fairness by ensuring parties retain the opportunity to challenge any information influencing an award.
California’s statutory action mirrors emerging judicial developments outside the United States. In ARIHQ v. Santé Québec, the Quebec Superior Court annulled an arbitral award over the improper delegation of decision-making authority to AI, according to the Mayer Brown analysis.
Together, these rulings suggest that enforceable legal boundaries are rapidly replacing reliance on voluntary guidance alone. Yet the California statute arrives precisely as arbitration institutions experiment with greater automation.
Whether these operational safeguards satisfy strict statutory rules requiring human decision-making remains unresolved. Additional uncertainties involve defining what constitutes delegating the practice of law and determining whether court-focused disclosure rules extend directly to arbitral submissions.
Legal professionals and arbitrators must review internal AI practices before Jan. 1, documenting tool usage and confirming that confidential materials remain isolated in restricted systems. Parties should establish clear AI protocols at the outset of disputes and discuss expectations with tribunals early.
The stakes extend far beyond inaccurate briefs, as awards showing undisclosed AI involvement or improper delegation now face immediate legal challenges grounded in statute and procedural fairness. For practitioners, retaining professional responsibility requires demonstrable operational controls over how automated systems participate in legal work.
Disclaimer: The information provided in this article is for educational and informational purposes only and does not constitute financial advice.
