AI in court: California adopts SB574 on lawyers’ responsibility
Signed on 30 September 2026, SB574 addresses citations, confidential information and human judgment. What will change in California, and what can a Belgian law firm learn?
A signed law with an implementation date to prepare for
California has adopted SB574, now Chapter 858. The governor approved it on 30 September 2026, and the legislative portal published the text on 1 October. The law addresses, among other matters, generative AI use by lawyers and arbitrators.
The text has no urgency clause or special commencement date. Article IV, section 8, of the California Constitution points to an effective date of 1 January 2027. That date follows from our reading of the two texts together. Signature therefore does not mean the new provisions already apply.
Every filed citation must be personally checked
The amendment to section 128.7 of the Code of Civil Procedure requires the lawyer responsible for filing to have personally verified every citation. This includes AI-generated references, without restricting the requirement to those references.
The practical problem extends beyond an invented judgment. A source may exist but address a different issue or support the opposite proposition. We recommend reading the original text, checking the passage relied upon and retaining a record of that review before filing. A reassuring response from the same assistant does not replace this step.
Assistance remains possible; responsibility stays human
New section 6068.1 prohibits delegating the practice of law to generative AI. It requires reasonable accuracy checks and correction of errors or hallucinations in material used by the lawyer. It also requires disclosure of AI use to the court for documents submitted to it.
A firm therefore needs substantive review with an identified person responsible. For example, an automatically prepared summary can help assemble a file; the lawyer must still verify the sources, facts and arguments chosen for submission. This illustrates a working method, not a decided case.
Confidentiality also depends on system access
SB574 addresses confidential, identifying and nonpublic information entered into a system where access to that information is not restricted to the lawyer and authorised people under confidentiality obligations. It is not a blanket ban on every AI tool.
Before sending a file, our practical recommendation is to examine who can read the data, how it is retained and what the provider undertakes to protect. A commercial « private » button does not answer these questions. The State Bar already provides ethical resources on these uses; those resources remain distinct from the enacted legislation.
Arbitrators must retain decision-making
New section 1282.1 prohibits an arbitrator from delegating any part of decision-making to a generative AI tool. For generated information outside the record, it provides for advance disclosure to the parties and, as far as practical, an opportunity to comment.
The distinction matters: practical help with preparation does not give a system the authority to decide. Parties can ask how information was obtained and check that the discussion remains grounded in the record.
What a Belgian reader should take away
SB574 is California legislation. It does not directly amend Belgian law or the EU AI Act. In an international matter, a professional must check the relevant jurisdiction, local rules and timetable before concluding that a duty applies.
When designing internal procedures, a firm can still use three questions: have the original references been read, may the data be sent to this tool, and who takes responsibility for final approval? This comparative analysis is not legal approval of an individual file. See our guide to evaluating a legal assistant and our editorial policy.
Sources and references
General information. Application to a matter depends on its facts and the rules in force.
