California SB 574 Survival Guide: What Every Litigator Must Change Before January 1, 2027
A practical breakdown of enacted Business and Professions Code § 6068.1(a)(3)(A)–(C), confidentiality boundaries on cloud AI tools, and why pre-certified AI boilerplate violates the statute.
1. What Enacted SB 574 Actually Requires of California Counsel
Signed into law on September 29, 2026 (Stats. 2026, ch. 858) and taking effect January 1, 2027, Senate Bill 574 adds Business and Professions Code § 6068.1, establishing explicit statutory duties whenever a California licensee uses generative artificial intelligence in the practice of law.
Rather than banning generative AI in litigation, the Legislature codified three non-delegable duties inside § 6068.1(a)(3): protecting confidential and personally identifiable information from unauthorized AI training systems, independently verifying every legal citation and factual proposition before filing, and exercising independent professional judgment to correct biased or inaccurate outputs.
(A) Ensure confidential, personally identifiable, or other nonpublic information is not input into a generative AI system that lacks adequate confidentiality protections or uses inputs to train publicly available models; (B) Verify accuracy of all AI-generated material, including confirming legal citations refer to existing law and support the proposition cited; (C) Exercise independent professional judgment and review outputs for bias and inaccuracy.
2. Why Enterprise Cloud Terms (Vertex AI CDPA) Matter More Than Marketing Claims
Under § 6068.1(a)(3)(A) and § 6068.1(b)(2), uploading raw arrest reports, medical/rehabilitation history, Social Security numbers, or driver's license identifiers to consumer AI endpoints that retain prompts for model training creates immediate disciplinary exposure alongside Business and Professions Code § 6068(e)(1) and Rule of Professional Conduct 1.6.
Even when using enterprise cloud infrastructure governed by a formal data processing addendum, where customer prompts and grounding queries are contractually excluded from model training, best practice remains client-side redaction before transmission. Stripping SSNs, dates of birth, minor identifiers, financial account numbers, and substance-treatment PHI into tracked [INSERT: REDACTED ...] placeholders before inference eliminates unnecessary exposure altogether.
3. Never Sign a Pre-Certified AI Verification Footnote
One of the fastest ways to turn a drafting error into an Order to Show Cause is signing a pleading containing pre-printed boilerplate asserting that 'all cited legal authorities have been independently verified' before you have actually opened the cases in an official reporter.
Remember also to match your compliance note to your forum: Code of Civil Procedure § 128.7(b)(2) applies to California civil and family law pleadings, whereas California criminal filings are governed by Business and Professions Code §§ 6068(d) and 6068.1(a)(3) and Rule of Professional Conduct 3.3(a)(1).
- Pre-scrub SSNs, DOBs, DL/CDL numbers, minor names/schools, financial accounts, and medical/rehab history locally before cloud drafting.
- Open every cited case, statute, and rule in CourtListener, Official California Appellate Reports, or California Legislative Information (LegInfo).
- Confirm pinpoint pages and verify that paraphrased holdings match the actual disposition, not an overruled or partially disapproved point.
- Remove all yellow [INSERT: ...] draft placeholders only after primary-source verification is 100% complete.
