The California AI lawyer bill would require attorneys to disclose generative AI use in documents submitted to a court and personally verify every legal citation. Lawmakers passed Senate Bill 574 unanimously on 31 August, but the measure is not yet law. It was enrolled on 4 September and still awaits the governor’s decision.
The bill goes beyond a warning about hallucinated cases. It says a lawyer must not delegate the practice of law to generative AI, must protect confidential and non-public information entered into a system, and must take reasonable steps to verify output and correct fabricated material. The human professional remains responsible for every filed document.
What California lawmakers approved
The Assembly passed SB 574 by 75 votes to none, and the Senate concurred by 40 to none, according to the Legislature’s official vote record. The absence of opposition gives the proposal political weight, but it does not replace the governor’s signature or allow the duties to be described as already in force.
For court documents, the bill uses a clear disclosure rule. A lawyer who used generative AI in preparing a filing would have to tell the court. The attorney would also have to personally verify every citation, including cases, statutes and other legal authorities, rather than trusting a generated reference or delegating the check back to software.
The measure also covers public-facing legal content and says lawyers should consider disclosure when AI materially contributed. It restricts confidential or non-public information from being entered into a system unless access is limited to authorised people. Those provisions turn familiar professional duties into explicit rules for an AI workflow.
The California AI lawyer bill reaches arbitrators too
SB 574 would stop arbitrators from delegating decision-making to generative AI. It would also require disclosure if an arbitrator used AI to obtain information outside the evidentiary record. That matters because private arbitration can determine employment, consumer and commercial disputes with less public visibility than a courtroom.
The rule does not ban administrative help or every use of AI. It preserves a distinction between supporting a professional and replacing the judgement that the professional was appointed to exercise. An arbitrator may use tools within the permitted boundaries, but cannot hand the actual decision to a model.
That principle resembles the concern raised in LiveAIWire’s analysis of AI evidence in court. Legal decisions depend not only on whether information is accurate, but whether its source can be challenged and the decision-maker can explain how it affected the result.
What the bill means for clients and lawyers
Clients should ask whether a firm uses generative AI, what information may enter the tool and who checks the result. Disclosure to a court does not automatically explain a firm’s full workflow to its client. Engagement letters and internal policies should state which systems are approved, whether client data are retained and how human verification is recorded.
Lawyers do not need to abandon useful tools while the bill is pending. They need a process that can survive scrutiny: approved platforms, access controls, source-grounded research and a named person responsible for checking every proposition and citation. Saving the prompt is not a substitute for opening the reported case and confirming that it says what the filing claims.
People representing themselves are not given the same professional infrastructure. LiveAIWire’s report on AI hallucinated citations in court found that self-represented litigants account for a large share of logged incidents. Court self-help services will still need practical guidance for people who cannot afford database-grounded legal tools.
California is formalising duties that already exist
The State Bar of California has already told lawyers that professional competence, confidentiality, candour and supervision apply when generative AI is used. Its updated practical guidance says lawyers must understand relevant benefits and risks, review generated output and avoid exposing protected client information.
SB 574 would give those expectations a clearer statutory form. The shift matters because guidance advises, while legislation can create an explicit compliance obligation. It also standardises disclosure instead of leaving every court or judge to invent a separate certification requirement.
California’s Judicial Council has adopted related rules for judicial officers and court staff. The court rule on generative AI requires policies that protect confidential information and ensure human review. The new bill addresses lawyers and arbitrators appearing alongside that institutional framework.
The disclosure rule is broader than catching fake cases
Fabricated citations created the most visible legal-AI scandals because they are easy to demonstrate. A case either exists or it does not. More subtle errors include misstating a real holding, omitting adverse authority, blending law from different jurisdictions or producing a confident argument that ignores facts unfavourable to the client.
Personal verification therefore requires more than searching for a case name. The lawyer must read the source, confirm the quotation, check that the authority remains good law and decide whether it supports the proposition in context. AI can accelerate discovery of possible authorities, but professional judgement begins where the generated answer ends.
Confidentiality creates a separate failure mode. A perfectly accurate output can still breach duties if a lawyer obtained it by pasting privileged facts into an unsuitable public system. LiveAIWire’s guide to AI use in divorce court explained how chatbot conversations can create discovery and privilege risks for clients as well as counsel.
Disclosure could expose differences between courts
A statewide rule would reduce uncertainty for California practitioners, but national litigation still crosses jurisdictions with different local orders. Some judges require a certification, some ask only for verification, and others rely on existing conduct rules. A firm cannot assume that complying with one California filing satisfies every court.
Standardised disclosure may also generate useful evidence about adoption. Courts currently know about AI mostly when something fails. Routine notices could show how often tools contribute to filings that cause no problem, helping policymakers distinguish ordinary assisted drafting from reckless reliance.
There is a risk that a generic disclosure becomes meaningless boilerplate. A notice should identify material use clearly enough for the court to understand what needs scrutiny without forcing lawyers to reveal privileged strategy. Implementing rules and judicial practice will determine whether the requirement improves accountability or merely adds another sentence to a signature block.
The bill is passed, but not signed
The official bill status lists SB 574 as active, passed and enrolled. That means the Legislature has finished voting and prepared the final text for the governor. It does not mean the proposal has taken effect, and coverage should be updated if it is signed, vetoed or allowed to become law through the applicable process.
The timing explains why the legislative vote drew attention. California’s legal market is large, and its approach can influence professional rules elsewhere. Unanimous approval also suggests disclosure and verification are becoming baseline duties rather than partisan limits on technology.
Still, the law would govern professional conduct, not make AI output reliable. A disclosure does not correct an invented authority, and a signed verification can fail if performed carelessly. Enforcement, training and firm-level controls will determine whether the rule changes behaviour.
AI can draft, but it cannot carry the duty
SB 574 draws a line that legal practice has been approaching for several years. A lawyer may use a model to organise, compare or draft, but cannot transfer the professional obligation attached to the finished work. The person signing the document must know the record, verify the law and protect the client.
That is a more durable response than banning a particular product. Models and vendors will change, while duties of competence, confidentiality and candour remain. Technology can alter how the work is produced without changing who answers when the work is wrong.
California has voted to make that responsibility visible in court. The next step belongs to the governor, and then, if the bill becomes law, to the lawyers and judges who must turn disclosure from a formality into a reliable record of human accountability.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.