What was signed

On 30 September 2026 Governor Gavin Newsom signed Senate Bill 574, introduced by Senator Thomas Umberg. It was one of thirteen bills listed in a single release from the Governor's office on AI, and was filed with the Secretary of State the same day as Chapter 858, Statutes of 2026. Most of the package concerns workers, healthcare and synthetic content. SB 574 is the one addressed to the legal system itself: attorneys, arbitrators, judicial officers and providers of alternative dispute resolution.

The release summarises it as keeping lawyers responsible by prohibiting them from fully handing over core legal work, such as drafting briefs or providing legal judgment, to AI. The statute is both more precise and wider than that. It contains a confidentiality rule, a verification duty, a duty of disclosure to courts, a citation rule backed by the existing sanctions procedure, and two rules for arbitrators.

The legislative history records no votes against it: 39–0 in the Senate in January 2026, 75–0 in the Assembly on 31 August and 40–0 in the Senate on concurrence the same day. It has no urgency clause, so under the ordinary constitutional rule for statutes passed in a regular session it takes effect on 1 January 2027. Bloomberg Law's report on the signing describes it as a first-of-its-kind bill. That is Bloomberg's characterisation; for our part, we are not aware of an earlier US statute regulating lawyers' and arbitrators' use of generative AI in these terms. Courts, bar bodies and arbitral organisations had issued rules and guidance before; this is legislation.

What the statute says

The chaptered text amends two codes in five places.

Business and Professions Code § 6068.1 (new): attorneys

The new section sits in the State Bar Act, immediately after § 6068, which lists an attorney's duties. It preserves the duty of reasonable competence and diligence and then states, as a free-standing rule, that an attorney shall not delegate the practice of law to generative artificial intelligence. An attorney who uses generative AI to assist in the practice of law must:

  • Protect confidentiality. 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 authorised who are under obligations to protect its confidentiality.
  • Verify and correct. Take reasonable steps to verify the accuracy of AI outputs, "including, but not limited to, the accuracy of all case and statutory citations", and to correct any erroneous or hallucinated output in any material the attorney uses.
  • Disclose. Disclose the use of generative AI to the court for all documents submitted to the court, and consider whether to disclose it where it is used to create content provided to the public.

"Personal identifying information" is defined by a list that includes dates of birth, social security and driver's licence numbers, addresses and phone numbers of parties, victims, witnesses and court personnel, medical and financial information, and anything sealed or deemed confidential by court rule or statute.

Code of Civil Procedure § 128.7 (amended): citations and sanctions

Section 128.7 is California's counterpart to Federal Rule 11. SB 574 adds that a brief, pleading, motion or any 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.

Because the rule sits inside subdivision (b), the section's existing sanctions apply. A sanction may follow a motion or the court's own order to show cause, but in either case the paper can be withdrawn or corrected within 21 days. Sanctions are limited to what deters repetition. Absent exceptional circumstances, the law firm is jointly responsible for violations by its partners, associates and employees. As the amended text reads, monetary sanctions may not be awarded against a represented party for a breach of the citation paragraph, so the money exposure falls on counsel and the firm rather than the client.

Code of Civil Procedure § 1282.1 (new): arbitrators

In the California Arbitration Act, after preserving the arbitrator's duties of neutrality, competence and diligence, the new section provides that an arbitrator shall not delegate any part of their decisionmaking process to any generative AI tool. An arbitrator also may not rely on AI-generated information outside the record without disclosing it to the parties beforehand and, as far as practical, allowing them to comment. The rule applies to "an arbitrator", not only to neutral arbitrators.

The other two changes

New § 180 of the Code of Civil Procedure requires the Judicial Council to publicly revisit, and revise as necessary, Standard 10.80 of the Standards of Judicial Administration. Since 1 September 2025 that standard has guided judges' use of generative AI in adjudicative work. Amended § 6173 of the Business and Professions Code requires the State Bar's voluntary ADR certification programme to include confidential procedures for complaints against certified providers. After a decision, the State Bar may publish the name of a provider found in breach.

Where the rules came from

Very little in SB 574 is new in substance. What matters is that it turns judicial warnings and soft-law texts into statute, and the wording shows its sources.

The citation rule comes almost word for word from Noland v. Land of the Free, L.P., decided by the Second Appellate District on 12 September 2025. The court found that 21 of the 23 case quotations in the appellant's opening brief had been fabricated by the AI tools counsel used. It published the opinion "as a warning", sanctioned counsel USD 10,000 and directed that the opinion be served on the State Bar. It said no filed paper "should contain any citations" that the responsible attorney "has not personally read and verified". The Legislature turned "should" into "shall", and kept "personally verified" while dropping "read".

The confidentiality limb, the list of personal identifying information and the formula on content "provided to the public" are taken from Standard 10.80, which already applied them to judges. SB 574 extends them to attorneys and adds a mandatory duty of disclosure to the court that the judges' standard does not contain.

The arbitrator provisions draw on Guidelines 6 and 7 of the Silicon Valley Arbitration & Mediation Center's AI Guidelines (April 2024). The non-delegation rule adapts Guideline 6, and the rule on information outside the record reproduces Guideline 7 almost word for word. A text designed for adoption in procedural orders is now California law.

The State Bar has moved in the same direction. Its Practical Guidance on generative AI dates from November 2023, and on 14 May 2026 its Board of Trustees approved a revised version that addresses agentic AI. All of this follows the run of fabricated-citation cases that this site examined in its 17 September analysis of the AI hallucination cases. Until now, judges and professional bodies had written that story. With SB 574, a legislature has joined it.

Legal analysis: what the text settles and what it leaves open

"Delegate the practice of law" is not defined. The saving clause on competence and the detailed duties that follow suggest the target is handing over judgment, not drafting with a tool. Until the State Bar or a court interprets it, the boundary will be drawn case by case.

Two verification standards sit side by side. Section 6068.1 requires reasonable steps to verify AI output. Section 128.7 requires every citation in a court filing to have been personally verified by the responsible attorney, whatever its source. The second is stricter and has no reasonableness qualifier. It also covers a citation supplied by a junior colleague or a template just as much as one produced by a chatbot.

Confidentiality turns on access, not on brand. The test is whether access to inputs is restricted to the attorney and authorised persons bound by confidentiality. That is a question about the provider's contract and architecture. An enterprise service with contractual confidentiality may pass. A consumer account whose terms let the provider review or train on inputs is unlikely to. The text contains no exception for client consent.

The disclosure duty is broad and has no stated form. All documents submitted to the court in which generative AI was used must be disclosed. The statute does not say how, in what detail, or whether incidental uses such as AI-assisted translation count. The definition of generative AI is wide enough to capture features built into ordinary office software. Courts will have to settle the practice.

The sanction route is conventional. Section 6068.1 states no penalty of its own. Given where it sits, breaches will naturally be handled through State Bar discipline. The citation rule uses § 128.7, including the 21-day safe harbour, which rewards prompt correction. One drafting point deserves attention. Section 128.7(d)(1) bars monetary sanctions against represented parties for breach of "paragraph (2) of subdivision (b)". Before the amendment, that paragraph dealt with legal contentions. After the renumbering it is the new citation rule. Whether or not this was intended, the text now shields clients from money sanctions for citations, and no longer expressly does so for legal contentions.

The arbitrator rules raise questions of reach and remedy. Section 1282.1 sits in the California Arbitration Act. The text does not address whether it reaches arbitrations under California's separate title on international commercial arbitration or the Federal Arbitration Act. Nor does it say what follows a breach. Losing parties can be expected to invoke it against awards. How courts fit such arguments into the existing grounds for vacatur remains to be seen.

How Türkiye compares

Türkiye has no statute on lawyers' use of AI, but the underlying duties exist in general form. Article 34 of the Attorneys' Act (No. 1136) requires lawyers to perform their work with care, accuracy and honour and to follow the professional rules of the Union of Turkish Bar Associations (TBB). Article 36 prohibits disclosure of matters learned in practice. Under the Code of Civil Procedure (No. 6100), Article 29 requires parties to act in good faith and state facts truthfully. Article 329(2) allows a disciplinary fine against a party who defends in bad faith or sues without any right, and directs it at counsel where counsel caused that conduct. Neither provision fits an unverified citation neatly.

There is also a structural difference. Under Article 33 of the same Code the judge applies Turkish law of their own motion, so a wrong citation does not bind the court as it can mislead an adversarial court that relies on counsel's research. The harm is real but different: lost time, lost confidence in counsel, and a sound argument discredited by a false reference.

The professional guidance is now close to California's. On 28 August 2026 the TBB published its Recommendation Guide on the Use of AI by Lawyers. It says a lawyer may not submit any provision, judgment, citation or fact obtained from AI to a court, arbitrator, mediator or authority without confirming it from an official and reliable source. It adds that a hallucinated citation may, depending on the facts, be a serious breach of the duty of care. Non-anonymised client data should not go into public tools or tools with inadequate contractual safeguards. On disclosure the guide is narrower than SB 574. The lawyer discloses AI use where the court or applicable procedure requires it, and otherwise remains personally responsible for accuracy. The guide is a recommendation, not a binding rule. On 22 September KVKK added a data protection layer, reported on this site: submitting personal data to a foreign-hosted AI service may be a transfer abroad requiring an Article 9 basis.

Enforcement has been more abrupt. In a matter reported in June 2026, the lawyer concerned gave this account: a civil court of first instance in Kızılcahamam, Ankara could not find on UYAP, in the form given, the Court of Cassation decisions cited in a defence petition. That same day it wrote to the Ankara Bar Association and the public prosecutor. The lawyer says the quoted content came from genuine judgments but the AI tool had attached the wrong chamber and docket numbers. The lawyer also says no specific offence has been identified. A lawyers' rights group criticised the use of criminal process, and the outcome is not known. California routes the problem into civil procedure, with a correction window and firm responsibility. In the Turkish matter, as reported, the court went straight to a bar and criminal referral.

How Korea compares

Korea's Attorney-at-Law Act already contains the two key duties. Article 24(2) forbids a lawyer to conceal the truth or make false statements in their work, and Article 26 imposes confidentiality. There is no AI-specific statute.

The courts have moved quickly. On 11 June 2026 the Patent Court stated in an invalidation judgment that many of the precedents the plaintiff relied on did not exist and that submitting them unverified was inappropriate. Counsel had cited eight AI-generated precedents in a motion to reopen argument, as Law Times reported. The Suwon and Seoul Eastern District Courts made similar findings in May and June, in cases brought by unrepresented litigants.

These rulings followed a National Court Administration task force of eight judges and two lawyers, whose results were announced on 31 March 2026. Courts may make a party or representative bear costs caused by AI-fabricated authorities and record the falsity in the judgment. At the panel's discretion, they may also refer the lawyer to the Korean Bar Association for discipline. The task force also proposed amending the Civil Procedure Rules to require parties to disclose AI use and verify key content. That would be the nearest Korean counterpart to SB 574's disclosure duty, but at the level of court rules. We could not confirm whether it has been adopted. The Seoul Bar Association's AI committee proposed member guidelines in December 2024. We could not verify a formal AI-use guideline from the Korean Bar Association.

What law firms should do

  • Make citation checking a recorded step. Every authority in a filing should be checked against an official or reliable source by the lawyer who signs, whatever its origin, and the check should be noted on file.
  • Classify tools by access, not by name. For each AI service, establish who can access inputs, whether they are retained or used for training, and what confidentiality obligations bind the provider. Only tools that pass should receive client information.
  • Prepare a disclosure practice. For California filings, decide in advance how AI use will be described and who signs off. Elsewhere, expect a court to ask.
  • Write down where assistance ends. A short policy on which tasks may be AI-assisted and which require independent judgment is the best evidence that the practice of law has not been delegated.
  • Plan for correction. Both California's 21-day window and the TBB guide reward a firm that can withdraw a defective filing quickly.

What arbitral institutions and arbitrators should do

  • Raise AI at the first procedural conference. The CIArb Guideline (updated September 2025) encourages arbitrators to consult the parties on their own use of AI and to refrain from a tool the parties disagree on. A short clause in the procedural order settles expectations and protects the award.
  • Keep the reasoning visibly human. Tribunals should be able to show that the analysis of evidence and law was their own, and that any AI-generated material from outside the record was disclosed with an opportunity to comment.
  • Publish an institutional position. Institutions such as ISTAC and the KCAB can expect parties and counsel familiar with the California rules to bring the same expectations. A published position is better than one improvised under challenge.

What to watch

  • How California courts implement the disclosure duty from 1 January 2027: local rules, forms or standing orders.
  • The Judicial Council's revision of Standard 10.80, and whether judges receive disclosure duties comparable to those now imposed on lawyers.
  • The first disciplinary or appellate interpretation of "delegate the practice of law", and the first challenge to an award under § 1282.1.
  • In Türkiye, the outcome of the Kızılcahamam referral and whether the TBB's recommendations reach the professional rules; in Korea, the fate of the disclosure proposal.

Frequently asked questions

Does SB 574 ban California lawyers from using AI?

No. It prohibits delegating the practice of law to generative AI and attaches conditions to its use: confidentiality, verification and correction, and disclosure to courts.

When does it apply?

It was approved and chaptered on 30 September 2026 without an urgency clause, so under the general rule it takes effect on 1 January 2027.

Does it affect lawyers outside California?

Section 6068.1 applies to attorneys regulated by the State Bar of California, and the citation rule to papers filed in California courts. Foreign lawyers will meet these rules in practice through California counsel and California filings.

Can an arbitrator use AI at all?

Yes, but not for any part of the decision-making process. Reliance on AI-generated information from outside the record must be disclosed in advance and, as far as practical, the parties must be able to comment.

Burhan Doğuş Ayparlar's View

This section sets out my personal assessment as the founder of this site and an attorney at law (Türkiye).

What strikes me most about SB 574 is how little of it is new. The citation sentence comes from a 2025 appellate opinion, the confidentiality and disclosure language from a judicial standard, and the arbitrator rules from a private soft-law text. The Legislature took what courts and practitioners had already concluded and gave it the force of statute. I think that is the right order of events. Rules for a new technology are best written after the profession has learned, at some cost, what goes wrong.

The citation rule persuades me most, because it is not really about AI. It says the lawyer who signs must have personally verified every authority, whatever its source. That was always the duty; AI simply made its breach visible in volume. A rule framed around the signer rather than the tool will still make sense when today's tools are obsolete.

I am less comfortable with the disclosure duty. Disclosing every document in which generative AI was used, with no threshold and no form, risks a ritual sentence at the foot of every filing. Such a sentence tells the judge nothing about whether the content was checked, which is the only thing that matters. I would rather see the duty attach to what the lawyer did to verify the output than to the fact that a tool was opened.

For Türkiye, the useful lesson concerns procedure. The TBB's August guide already states the verification duty as clearly as California does. Its more modest line on disclosure is, I think, defensible. What we lack is a proportionate route between a mistaken citation and a criminal referral. The Kızılcahamam matter shows a court facing a new problem without a tailored tool and reaching for the heaviest one available. California's model separates carelessness from deceit. It puts sanctions in civil procedure, allows time to correct, holds the firm responsible and leaves discipline to the bar. A lawyer who files an unverified citation has failed in a professional duty under Article 34 and should answer for it before the bar. Whether that is also a crime is a different question, and our procedure should not treat the two as one.

Finally, the arbitration provisions deserve more attention than they will get. Arbitration's legitimacy rests on the parties' choice of a person to decide their dispute, and an award reasoned in substance by a model breaks that bargain whatever any statute says. Institutions in Istanbul and Seoul do not need California's legislature to tell them so. They have every reason to say it in their own rules before a challenge forces the issue.

This article is for information only and does not constitute legal advice. It is based on the chaptered text of SB 574 (Chapter 858, Statutes of 2026) and its legislative history on leginfo.legislature.ca.gov, the Governor's release of 30 September 2026, Bloomberg Law's report on the signing, the Noland opinion, Standard 10.80, the SVAMC and CIArb guidelines, the TBB guide of 28 August 2026, and Turkish and Korean statutes and press reports available on 8 October 2026. The Kızılcahamam and Korean court matters are described from press reports; the underlying decisions were not available to us. Whether the Korean task force's proposals have been adopted, and whether the Korean Bar Association has issued a formal AI guideline, could not be verified. Statements about Turkish and Korean law are general in nature; specific cases require individual assessment. The analysis and assessments are the author's own.