For many courts and law offices, AI guidance still sounds like a list of good habits: protect confidential information, check the answer, keep a human responsible, and do not let a tool make the decision.
California has now put much of that list into a statute.
On September 30, California approved and chaptered SB 574 as Chapter 858, Statutes of 2026. Under California's ordinary constitutional effective-date rule, it takes effect January 1, 2027.
The statute does not ban generative AI. It does something more practical. It draws clear lines around responsibility.
For lawyers, the law says generative AI cannot be delegated the practice of law. A lawyer who uses it must protect specified confidential, personal identifying, and other nonpublic information. The lawyer must take reasonable steps to verify AI output, including case and statutory citations, and correct erroneous or hallucinated output used in the lawyer's work.
The filing rule is equally plain. A brief, pleading, motion, or other court paper cannot contain a citation the submitting attorney has not personally verified, including a citation supplied by generative AI.
The statute also reaches beyond lawyers. It bars an arbitrator from delegating any part of the decisionmaking process to generative AI. If an arbitrator relies on AI-generated information outside the record, the arbitrator must make appropriate disclosures to the parties and, as practical, allow comment. California's Judicial Council must also publicly revisit Standard 10.80 of the Standards of Judicial Administration and revise it as necessary for the further development of generative AI.
California's statute is not a Tennessee court rule, and it does not govern Tennessee lawyers or judges. But it gives every court and law office a useful operating model. Written AI policy becomes real only when it answers four questions.
Four questions before AI assists legal or court work
- Who remains responsible? A tool can assist. It cannot inherit a lawyer's duty, an arbitrator's neutrality, or a court's judgment.
- What information may enter the tool? Confidential case material, personal identifying information, sealed information, and other nonpublic records need a defined boundary before anyone starts copying material into a system.
- What must be checked before it matters? Citations are the obvious example, but the same discipline applies to quotations, factual summaries, translations, research results, and drafted communications.
- What must be disclosed or preserved? A court or law office should know when AI materially shaped a filing, an analysis, or a decision-support step. The answer may differ by jurisdiction and task, but it should not be improvised after an error.
The lesson is not that every jurisdiction should copy California. The lesson is that broad AI principles become useful when they are tied to a real work product, a named human owner, a data boundary, a review step, and a record.
Question for local discussion
Before AI assists with legal or court work, can the responsible person explain what the tool may do, what it may receive, what must be checked, and who remains accountable?