California Writes Human-in-the-Loop into Law

Generative AI has quickly become part of everyday legal work, and lawmakers are now setting rules for it. On September 30, California enacted SB 574, which sets requirements for how attorneys use generative AI. The law takes effect January 1, 2027. As BakerHostetler summarizes, it applies to California-licensed attorneys, attorneys appearing before California courts and arbitrators in California proceedings. It requires them to:

• Not delegate the practice of law to generative AI
• Take reasonable steps to verify AI-generated content and correct errors or hallucinations
• Protect confidential and personally identifying information
• Meet applicable disclosure obligations
• Personally verify citations in California court filings

None of this should surprise anyone who has followed how courts handle AI. SB 574 simply puts into statute what we’ve argued for some time: delegating work to AI is not the same as delegating responsibility. It also signals that additional regulatory requirements are likely on the horizon.

What the law codifies

In our recent post, Human-in-the-Loop and What Courts and Regulators Expect, we described how defensible AI workflows keep humans in the loop at the points where it matters most:

• Attorneys review prompts and early samples before AI classifies a full collection.
• Privilege QC includes sampling documents the AI marked non-privileged.
• Escalation rules define which decisions need human approval.
• Production QC is the final safeguard.

SB 574’s verification and non-delegation requirements map directly onto these practices.

The rules aren’t new. The tools are.

Courts are already applying the same reasoning. In Schulte v. LinkedIn Corp. (N.D. Cal. June 30, 2026), the court approved LinkedIn’s plan to use Relativity aiR for responsiveness review, with human sampling and QC. The court applied the same proportionality, reasonableness, and transparency standards that govern TAR. Generative AI review did not need new rules. It needed the existing ones, applied with discipline.

What this means for legal teams

The lessons from SB 574 and Schulte is the same: use AI but be able to show your work. That means:

• Calibrated prompts with measurable accuracy outputs
• Statistical sampling methodologies to validate results, detect hallucinations and inaccuracies, and ensure source traceability and citation integrity.
• Documented quality control checkpoints and human decision points throughout AI and agentic workflows
• Robust data governance, access controls, privacy, and security requirements
• An attorney who stands behind every result, no delegation of legal judgment or the practice of law

Cimplifi has built its generative AI review workflows around this approach. Talk to our team about workflows that are efficient, defensible and ready for what California and the courts now expect.

About the Author
Matt Durney is an attorney, financial services specialist, eDiscovery project manager, and CLM consultant. Matt has worked with corporate financial services clients and their in-house and outside counsel on financial services litigation including both regulatory and commercial components. Matt speaks frequently on artificial intelligence, legal technology, and contract analytics.