The legal profession has spent the past few years asking whether generative AI is trustworthy. Will it draft a contract, research an issue, summarize a record, or invent a case with impressive confidence? California’s new law shifts the focus to a more practical question: who takes responsibility when lawyers use it?
On September 30, 2026, Governor Gavin Newsom signed Senate Bill 574, which legal industry news outlets described as a first-of-its-kind framework for attorneys’ and arbitrators’ use of generative AI. The law does not push AI out of legal practice. Instead, it sets ground rules for confidentiality, verification, disclosure, and human judgment.
What the Law Does
- No delegation of legal practice. An attorney cannot hand the practice of law to generative AI. The attorney must exercise professional judgment and remain responsible for the work.
- Accuracy and correction. An attorney who uses generative AI must take reasonable steps to check the output, including case and statutory citations, and fix errors or hallucinations before using the material.
- Disclosure in court filings. An attorney must disclose the use of generative AI to create a document submitted to a court.
- Personal verification of citations. An attorney responsible for a brief, pleading, motion, or other court filing must personally verify every citation, including citations supplied by generative AI.
- Protection of confidential information. An attorney cannot enter confidential, personally identifying, or other nonpublic information into an open-access generative AI system. The attorney may use a restricted system when only authorized people bound by confidentiality obligations have access.
- Rules for arbitrators. Arbitrators cannot delegate any part of the decision-making process to generative AI. The framework also addresses AI-generated information from outside the record, including disclosure to the parties and an opportunity to respond when practicable.
- Judicial and ADR oversight. The law calls for further attention to standards for judicial use of generative AI and oversight of certified alternative dispute resolution providers.
California has not placed generative AI off-limits. The law instead applies familiar professional duties to a tool that produces polished work without assuming responsibility for it. Lawyers have long relied on associates, paralegals, vendors, and technology. Generative AI presents a different challenge because the output often looks like legal judgment. The tool might draft an argument, find authority, flag litigation risk, or suggest a strategy. SB 574 draws the line at substitution: AI gets a seat at the worktable, not the lawyer’s chair.
Confidentiality puts that principle into everyday practice. Firms and legal departments need a clear list of approved tools, the safeguards attached to each tool, and the information users should keep out. “Be careful” is not much of an AI policy if nobody knows what careful means.
Disclosure also adds a checkpoint to the filing process. Firms should identify material AI involvement before a filing reaches final review, not five minutes before the deadline. A workable process should record that involvement, prompt attorneys to check the output and citations, and flag the need for disclosure.
The rule for arbitrators draws an even brighter line. AI might organize a record or sort information, but it does not get a vote. The decisive question is not merely whether the output is accurate. The question is whether the tool supports adjudication or begins to perform it.
Other regulated professions should take note. Healthcare, finance, insurance, employment, and corporate governance also involve AI-assisted decisions that a human professional must own. California’s approach offers a useful four-part playbook: protect the data, check the output, disclose meaningful use, and keep human judgment human.
Practical Considerations
Law firms, legal departments, arbitrators, and alternative dispute resolution providers should turn those principles into concrete procedures. Organizations should identify approved and open-access tools, define what information belongs in each system, and document the review of AI-generated work. Filing procedures should flag AI assistance, citation checks, and necessary disclosures. Training, supervision, and a clear escalation path should support those controls. The goal is straightforward: give AI useful work without giving it the keys to professional judgment.
The most interesting question is no longer whether AI is sophisticated enough to perform legal work. The harder question is which parts of professional judgment society will insist on keeping human—even when the machine gets very good.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
[View Source]