California lawyer AI rules put confidentiality and citation checks in focus

California has enacted rules addressing lawyers' use of generative AI, according to an October 1 analysis from Holland & Knight. The firm's alert says Governor Gavin Newsom signed Senate Bill 574 on September 30, with the requirements taking effect January 1, 2027.

The analysis describes duties involving confidentiality, verification of generated output and disclosure to courts. It also identifies questions the statutory language leaves open. This article reports that firm's analysis; it is not legal advice or an independent determination of how a court will interpret the law.

Attorney judgment remains the central boundary

Holland & Knight says new Business and Professions Code Section 6068.1 prohibits delegating the practice of law to generative AI while preserving duties of competence and diligence. The firm notes that the statute does not define delegation or specify an exhaustive set of permitted tasks.

Its interpretation allows AI assistance with activities such as research, drafting and document evaluation while leaving professional judgment with the attorney. That is an interpretation offered by the authors, not an explicit task-by-task exemption quoted from the law.

The distinction matters for automated workflows. A lawyer reviewing the reasoning and deciding what to use is different from approving an output without examining it. The alert argues that meaningful attorney involvement is important at substantive decision points, including in firms organized around extensive automation. It does not suggest that purchasing an AI product transfers professional responsibility to the vendor.

Confidential inputs require attention to access

The alert describes a prohibition on entering confidential, nonpublic or personally identifying information into generative AI unless access is restricted to the attorney and authorized people who have confidentiality obligations. It says this may exclude some publicly available consumer tools.

That makes a product's data-handling arrangements relevant to a lawyer's evaluation. The firm's analysis recommends examining the tool's features and the information in documents before uploading protected material. A reassuring product label alone does not answer the alert's access-restriction question.

The analysis also says attorneys must take reasonable steps to verify AI outputs, including case and statutory citations, and correct erroneous material they use. The authors describe checking whether a cited authority exists, remains good law and supports the proposition attached to it. Those checks address the substance of a citation as well as an invented case name.

Court citations and disclosure have separate provisions

Holland & Knight says an amendment to Code of Civil Procedure Section 128.7 requires the responsible attorney to personally verify every citation in a California court filing. The described requirement covers citations supplied by AI and those obtained from other sources.

The authors interpret personal verification as requiring more than sole reliance on someone else's check. They also note an unresolved question about which lawyer is responsible when several lawyers sign or appear on a paper. The alert places possible violations within Section 128.7's sanctions framework, rather than treating verification as an optional best practice.

Disclosure raises another uncertainty. The analysis says Section 6068.1 requires attorneys who use generative AI in their practice to disclose that use to the court for submitted documents. The authors discuss a document-specific reading and a broader possible reading, and say the statute does not prescribe the form of disclosure or the level of AI involvement that triggers it.

Their proposed cautious approach is disclosure when AI helped prepare a particular filing. That recommendation should remain attributed to the firm, especially while interpretation and implementation details are unsettled. Attorneys with a concrete matter need the enacted text, applicable court procedures and qualified advice.

The alert gives practitioners a set of issues to examine before January: permitted inputs, access controls, output and citation review, and court disclosure. Its most consequential point is that AI-assisted work still needs an accountable lawyer, with uncertainty about some procedural details left visible rather than replaced by a confident automated checklist.