California SB 574 (Chapter 858), Signed 30 September 2026, Sets Generative AI Duties for Lawyers, Court Filings and Arbitrators From 1 January 2027
On 22 June 2023 Judge P. Kevin Castel of the Southern District of New York sanctioned two lawyers and their firm in Mata v. Avianca. Their brief had cited judicial opinions that did not exist, produced by ChatGPT, and the lawyers kept standing by them after the court's orders called their existence into question. The penalty, under Rule 11, was USD 5,000 paid into the court's registry. The case became the standard anecdote of the generative AI era in law because every litigator understood how it happened: someone trusted a fluent answer and signed.
California has now turned that anecdote into statute. Senate Bill 574, authored by Senator Thomas Umberg, was approved by the Governor and filed with the Secretary of State on 30 September 2026 as Chapter 858 of the Statutes of 2026. It amends two codes, adds three sections and reaches lawyers, judges, arbitrators and the State Bar. There is no urgency clause, so under the California Constitution it takes effect on 1 January 2027. This post reads the chaptered text, section by section.
Practical takeaway. From 1 January 2027 a California lawyer who uses generative AI must keep client and personal data out of tools whose access is not restricted, verify every output and citation, disclose the use to the court for documents submitted to it, and never let the tool practice law. Arbitrators may not let it decide any part of a case.
What section 6068.1 of the Business and Professions Code requires of every California attorney
Section 1 of the bill adds section 6068.1 to the State Bar Act. It opens by preserving the existing duty of competence and diligence, then states the core rule in one line: "An attorney shall not delegate the practice of law to generative artificial intelligence." Use for assistance is permitted on three conditions.
First, the attorney may not enter confidential, personal identifying or other nonpublic information into a generative AI system unless access to what the attorney inputs is restricted to the attorney and persons the attorney has authorized, who are themselves under obligations to protect it. Second, the attorney must take reasonable steps to verify the accuracy of outputs, "including, but not limited to, the accuracy of all case and statutory citations," and to correct "any erroneous or hallucinated output" in material the attorney uses. The word "hallucinated" now appears in the California codes. Third, the attorney must disclose the use of generative AI to the court for all documents submitted to it, and must consider whether to disclose it when the tool helped create content provided to the public.
The statute defines generative AI broadly, as a system that can generate derived synthetic content, including text, images, video and audio, emulating the structure and characteristics of its training data. That definition covers drafting assistants, research tools with a generative layer, and image or audio tools used to prepare demonstrative exhibits.
Which data may not go into an unrestricted AI tool: medical, psychiatric and financial information on the statutory list
Subdivision (b)(2) of section 6068.1 lists what counts as personal identifying information. It names driver's license numbers, dates of birth, Social Security numbers, criminal identification numbers, the addresses and phone numbers of parties, victims, witnesses and court personnel, medical or psychiatric information, financial information, account numbers, and any content sealed by court order or made confidential by rule or statute.
That list carries the statute well beyond litigation technique. A medical malpractice file, a conservatorship petition, a family-law financial disclosure and a personal injury demand routinely contain medical, psychiatric or financial information. A patient whose records reach a plaintiff's firm, or a hospital's defense counsel, is now protected by a rule that turns on the AI vendor's access controls. The commercial consequence follows directly. The statute does not forbid a provider from ever seeing inputs: access is acceptable when it is limited to the attorney and persons the attorney has authorized who are bound to protect the information. What fails the test is access outside that circle, for example unrestricted provider use of inputs, or training that opens them to people outside the authorized confidential circle. Compliance depends on both the contractual restrictions, such as retention, training-use clauses and sub-processor lists, and the way the tool actually operates.
How amended Code of Civil Procedure section 128.7 changes what a signature on a California court filing certifies
Section 128.7 has governed signatures on California pleadings since the mid-1990s, and it applies to complaints and petitions filed on or after 1 January 1995. By signing, filing or later advocating a paper, an attorney certifies that it has a proper purpose and that its legal and factual contentions are warranted after reasonable inquiry. Chapter 858 adds a new paragraph (b)(2): a brief, pleading, motion or other paper filed in any court "shall not contain any citations that an attorney responsible for submitting the pleading has not personally verified, including any citation provided by generative artificial intelligence."
The rule is written for all citations, whatever their origin, in papers that section 128.7 covers. Subdivision (g) leaves its scope where it was: the section does not apply to disclosures and discovery requests, responses, objections and motions. Section 6068.1 is a separate professional duty, enforced by the State Bar, and its verification and disclosure requirements are not limited by that exception. A citation copied from an associate's draft, a treatise or an AI tool is treated the same way, and the verb is "personally". The existing sanctions machinery then applies unchanged: a 21-day safe harbor in which a challenged paper can be withdrawn or corrected, a court's power to issue an order to show cause on its own motion, sanctions limited to what deters repetition, and joint responsibility of the law firm for its partners, associates and employees absent exceptional circumstances. As the chaptered text reads, subdivision (d)(1) still bars monetary sanctions against a represented party for a violation of "paragraph (2) of subdivision (b)", which after the amendment is the citation rule. Within section 128.7, then, monetary sanctions for an unverified citation can reach the lawyers and the firm but not a represented party. That limit concerns this sanctions provision only; it does not protect a client from other consequences of a defective filing.
Federal practice reached a similar place by a different road. Federal Rule of Civil Procedure 11 carried the court in Mata, and many federal judges have since issued standing orders on AI-assisted filings. California has chosen statute over case-by-case orders, which gives every superior court the same rule on the same day.
What arbitrators may no longer do under new section 1282.1, and what changes for the State Bar's ADR certification
Section 5 of the bill adds section 1282.1 to the California Arbitration Act. An arbitrator "shall not delegate any part of their decisionmaking process to any generative artificial intelligence tool," and may not rely on AI-generated information outside the record without disclosing it to the parties beforehand and, as far as practical, letting them comment on its use. The rule sits beside the existing neutral-arbitrator disclosure duties and reaches the private forums where many employment, consumer and commercial disputes are decided.
Section 2 rewrites section 6173 of the Business and Professions Code, which governs the State Bar's voluntary program to certify alternative dispute resolution firms, providers and practitioners. The State Bar, or an approved professional organization, must now receive, investigate and resolve complaints that a certified provider failed to meet the applicable ethics standards. Proceedings are private and confidential, and the records are exempt from the Public Records Act. After a decision the State Bar may disclose information that reveals no confidential or privileged communication. The bill does not tie that complaint route to AI by its terms, but it gives a party who suspects an AI-written award an institutional place to raise it.
Section 4 adds section 180 to the Code of Civil Procedure and directs the Judicial Council to publicly revisit, and revise as necessary, Standard 10.80 of the California Standards of Judicial Administration, the standard on judicial officers' use of generative AI. Lawyers, arbitrators and judges are therefore all addressed in one chapter.
How SB 574 sits next to California's frontier-model law and the EU AI Act
California already regulates the companies that build the largest models through SB 53, whose reporting duties we tested against OpenAI's 20 September 2026 sandbox breakout. SB 574 works at the other end of the chain. It leaves model developers alone and regulates the licensed professionals who use the output, through the State Bar's discipline system and the courts' sanctions power. Between them the two statutes cover the builder and the professional user, with nothing yet addressed to the many vendors who package models into legal products.
The European Union approaches the same ground mainly from the system side. Annex III, point 8(a) of Regulation (EU) 2024/1689 lists AI systems intended to assist a judicial authority in researching and interpreting facts and law, and systems used in a similar way in alternative dispute resolution, as high-risk, subject to Article 6(3), under which such a system is not high-risk where it does not pose a significant risk of harm, including by not materially influencing the outcome of decision-making; narrow procedural and preparatory tasks are among the listed examples, and a system that profiles natural persons always stays high-risk. The AI Act also places duties on deployers, including human oversight and use in line with the provider's instructions. Its center of gravity is still the system and the organization that uses it. California's statute speaks to the individual arbitrator and forbids delegation of any part of the decision outright, whatever the classification of the tool.
How Quentir Reads It
The center of gravity in Chapter 858 is professional responsibility. The statute assumes lawyers will use generative AI and spends its words on where client data goes, who checks the citations and who answers when the tool is wrong. That choice makes the rule enforceable with institutions California already has, the State Bar and the courts, and it moves real attention onto contract terms: the access-restriction test in section 6068.1(a)(3)(A) will be satisfied or failed in a vendor's data-processing terms long before any disciplinary hearing.
The disclosure duty is the provision with the widest reach and the least detail. It requires disclosure to the court for all documents submitted, without prescribing a form, and asks lawyers to consider disclosure for public-facing content. Disclosure of AI-generated content is also the subject of Article 50 of the EU AI Act, and our Signature Brief 2026.3 on AI Act Article 50 disclosure reads that European obligation in depth, with fixed scope, dates and the questions to ask, which this post does not attempt. For California, the open question for 2027 is practical: whether the Judicial Council, in revisiting Standard 10.80 or through a rule of court, sets a uniform form for the disclosure, or whether each court will build its own.
Sources: California Legislature, Senate Bill 574 (Umberg), Chapter 858, Statutes of 2026, approved by the Governor and filed with the Secretary of State on 30 September 2026 (chaptered text published 1 October 2026); Judicial Council of California, Standard 10.80, Use of generative artificial intelligence by judicial officers; U.S. District Court for the Southern District of New York, Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), opinion and order on sanctions, 22 June 2023; Regulation (EU) 2024/1689 (AI Act), Annex III point 8(a) and Article 50. All sources accessed 3 October 2026.
Published intelligence, built to inform your own decisions. Published: October 3, 2026.