California's SB 574 rules on lawyers' use of AI are now law: on September 30, 2026, Governor Gavin Newsom signed the bill (Chapter 858 of the Statutes of 2026), and it takes effect on January 1, 2027. California thereby becomes the first US state to regulate attorneys' use of generative AI by statute — not through a judge's standing order or a bar association opinion, but through a new Section 6068.1 of the Business and Professions Code and amendments to Section 128.7 of the Code of Civil Procedure. Reuters and Bloomberg Law both reported the signing on October 1.
The three things this law actually governs
First, an attorney may not delegate the practice of law to generative AI. AI can help with drafting, research and organizing, but legal judgment and responsibility for the work product must stay with a licensed attorney. Sending AI-generated advice to a client without meaningful review, or filing AI-written papers as they came out, is exactly the use this rule targets.
Second, confidential information cannot simply be fed into any AI tool. Attorneys must not enter confidential, personally identifying or other nonpublic information into a generative AI system unless access to those inputs is restricted to the attorney and people the attorney has authorized who are under a duty to keep the information confidential. A consumer tool that retains inputs, allows vendor staff to review them, or trains on them will generally fail that test for client material. Enterprise tools are not automatically safe either: firms have to read the actual data-retention and human-review terms, and merely turning off training may not be enough. The statutory definition of personally identifying information is broad, covering driver's license and Social Security numbers, dates of birth, addresses and phone numbers of parties and witnesses, medical and financial information, account numbers, and information sealed by law.
Third, AI output must be verified and corrected, and court filings must disclose AI use. Attorneys must take reasonable steps to verify the accuracy of AI output and correct any erroneous or hallucinated material they use — a duty that covers not only citations but also factual summaries, deposition digests, contract analysis and any other material an attorney actually relies on. On top of that, any document submitted to a court that was produced with generative AI must disclose that use. That disclosure duty is the genuinely new daily obligation compared with earlier ethics guidance.
A personal duty the signing attorney cannot hand off
As amended, Section 128.7 of the Code of Civil Procedure adds a specific rule: a brief, pleading, motion or any other court paper must not contain citations that the attorney responsible for submitting it has not personally verified, including citations provided by generative AI. Two points matter. "Personally" means that leaving the checking entirely to a paralegal, a junior associate or a citation tool, without the responsible attorney confirming the citations, may not satisfy the duty. And the scope is not limited to AI citations: even a citation that did not come from AI must be personally verified by the signing attorney. The burden falls hardest on solo practitioners and small firms; large firms can spread checking work across junior lawyers, while a solo attorney has to verify every citation alone.
Arbitrators are covered by the same statute: they may not delegate any part of their decision-making to a generative AI tool, and if they rely on AI-generated information outside the record, they must first disclose that use to the parties and, as far as practicable, give them a chance to comment. The Judicial Council is also directed to revisit its judicial administration standards in light of generative AI.
Why now: fictitious citations have already drawn real sanctions
The law responds to two years of sanctions across US courts against lawyers who filed fictitious citations. In March this year, a California appellate court sanctioned counsel $5,000 for fictitious AI-generated citations in a custody dispute. Until now, states relied mainly on judge-by-judge orders and bar ethics opinions, which varied widely. SB 574 gathers those scattered expectations into one statute and attaches sanctions to the personal-verification duty: monetary sanctions for a citation violation cannot be imposed on a represented party, so the exposure falls on attorneys and their firms, and a firm is in principle jointly responsible for violations by its partners and employees. The statute keeps a 21-day safe harbor — withdrawing or correcting the challenged paper within 21 days of service of a sanctions motion can avoid sanctions.
In the same batch of California AI bills signed on September 30, another measure governs automated decisions in the workplace, and the two read well side by side: after California's SB 947, an automated system cannot be the sole basis for discipline or termination. One law governs how employers use AI on workers, the other how lawyers use AI in their cases; neither bans the tools — both ban handing judgment over to them.
What firms should do before January 1, 2027
Scope first: the statute governs attorneys practicing in California and filing papers in California courts, and out-of-state counsel appearing pro hac vice are very likely covered as well. It does not directly govern courts in other states, and it does not ban lawyers from using AI. The confidentiality and verification duties are not limited to court filings — client advice and internal memos are covered too.
Before the effective date, firms with California practice should do at least four things: review the data-retention, human-review and training terms of every AI tool in use, approve only tools that meet the access-restriction test, and expressly prohibit consumer tools for client information; write down what information may be entered into which tool, using the statute's definition of personally identifying information as the floor; build the signing attorney's personal citation check into the pre-filing checklist and document it; and adopt standard disclosure language for California filings, with a named person responsible for confirming whether AI was used on each document. The hardest question is the boundary of disclosure: generative AI is now built into research and word-processing tools, and the statute does not say which uses count or whether light editing qualifies. Until courts or the State Bar give guidance, preparing on a "material use means disclosure" basis is the lower-risk default. Treating AI as a shortcut that carries no responsibility, by contrast, becomes a clear statutory risk in California from next year, not merely an ethics reminder.
