California SB 574: what changes on January 1, 2027

Originally published on the MindHYVE Insights Substack on October 18, 2026, as an advisory for California law firms. Republished on MindHYVE.ai with light canon edits.

This advisory provides information about a new California statute and State Bar guidance. It is not legal advice and does not create an attorney-client relationship. Verify the statutory text against the chaptered bill before relying on it.

California has enacted the first state law in the country governing attorneys' use of generative AI.

SB 574 was signed on September 30, 2026 (Stats. 2026, ch. 858). With no separate operative date, it takes effect January 1, 2027.

The law does not ban AI in legal practice.

It does something more consequential: it converts professional-responsibility expectations into statutory duties and attaches an existing sanctions regime to one of them.

Most firms already believe they comply.

In our reading, most do not — not because of substantive misconduct, but because two new duties are workflow duties, and existing filing processes do not automatically perform them.

New: Business & Professions Code § 6068.1

The central provision states:

“An attorney shall not delegate the practice of law to generative artificial intelligence.”

Four affirmative duties follow:

• Confidentiality — Keep nonpublic client information out of systems not restricted to the attorney and persons under confidentiality obligations.
• Verification — Verify the accuracy of generative AI outputs, including all case and statutory citations.
• Correction — Correct hallucinated AI output.
• Court Disclosure — Disclose the use of generative AI to the court for all documents submitted to the court.

Amended: Code of Civil Procedure § 128.7(b)(2)

Filed papers must not contain citations the responsible attorney has not personally verified, including citations generated by AI.

Existing § 128.7 sanctions provisions apply.

Pay particular attention to the word PERSONALLY.

An automated citation checker is useful. It is good practice and potentially valuable evidence.

But it does not replace the signing attorney's own verification.

Any vendor suggesting otherwise is exposing attorneys to sanctions risk.

New: CCP § 1282.1

Arbitrators are similarly prohibited from delegating decision-making to generative AI.

The State Bar's COPRAC Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law was revised May 14, 2026, at the California Supreme Court's request to address agentic systems.

The guidance states that lawyers must not allow agentic systems to make substantive legal determinations, provide legal advice, prepare and file pleadings, or act in a representative capacity without meaningful lawyer supervision and review.

It also prohibits autonomous filing, court communications, and representations on the lawyer's behalf.

The distinction:

• ✅ The system drafts. The attorney reviews and files. ❌ The system autonomously files.
• ✅ The system organizes, extracts, and flags. The attorney evaluates the information. ❌ The system independently corresponds with the court.
• ✅ The system proposes. The attorney decides. ❌ The system makes representations on the attorney's behalf.

Practical consequence:

Autonomous filing and correspondence are not conveniences that can simply be enabled with loose supervision.

They are among the capabilities most likely to create conflict with California's agentic AI guidance.

Do not confuse proposed professional-conduct rule amendments with adopted rules.

Proposed amendments to the Comments on Rules 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3 remain proposed, not adopted, according to the cited October 1, 2026 status.

The second comment period closed August 6, 2026.

Adoption requires both Board of Trustees action and California Supreme Court approval.

Under Rule 1.0(c), Comments are not themselves a basis for imposing discipline.

Important sourcing caution:

The State Bar republishes revisions of its Practical Guidance using the same filename, without a revision date on the document's face or a changelog.

A copy downloaded last year may contain materially different guidance.

Verify the substance, not merely the URL.

Complete these before January 1, 2027.

1. Make court disclosure a standard template field

The statute's disclosure duty covers all documents submitted to the court.

Do not rely on individual attorneys remembering to disclose AI use.

Build disclosure into filing templates and pre-filing checklists.

2. Document personal citation verification

Introduce a source-by-source sign-off before the responsible attorney signs.

CCP § 128.7 carries sanctions exposure.

In a sanctions hearing, the issue may be what verification the attorney can demonstrate.

An undocumented verification is difficult to distinguish from no verification.

3. Audit client information access

Confidentiality involves access, not merely data retention.

Review:

• Vendor personnel with production access
• Shared workspaces
• Team accounts
• Third-party systems
• Contractual confidentiality obligations

Ask every AI vendor in writing: who can access client matter content, and under what confidentiality obligations?

4. Identify autonomous filing and communications

Inventory your AI tools against the COPRAC guidance.

Prioritize disabling features that autonomously:

• File court documents
• Contact opposing counsel
• Communicate with courts
• Make representations on behalf of attorneys

5. Update engagement letters and billing practices

Explain where AI is used.

Bill appropriately for attorney review and professional judgment, rather than treating machine-generated output as equivalent to attorney work.

Billing practices may present substantial downstream professional-responsibility risk.

6. Supervise — and document supervision

Meaningful lawyer supervision and review is the governing principle.

Rules 5.1 and 5.3 already impose supervisory responsibilities.

Make each review a documented event containing:

• Attorney identity
• Timestamp
• Matter reference
• Review or attestation status

Supervision should be demonstrable, not merely asserted.

JustineAI™ is MindHYVE's Agentic Legal Operating System, powered by Justine, our legal Digital Employee.

Two editions are currently in market:

• Personal Injury (PI)
• Workers' Compensation (WC) — including California applicant-side practice.

We did not build JustineAI for SB 574.

We built it around the same underlying principle that the statute now expresses:

“The AI reasons. The attorney decides.”

This principle is reflected in the system's architecture, workflows, and governance model.

A. Non-delegation and attorney control

Under § 6068.1 and the COPRAC guidance, practicing law and making substantive legal decisions remain attorney responsibilities.

JustineAI's stated architecture preserves that boundary.

Justine can: draft documents, organize matters, sequence workflows, identify issues, propose courses of action.

Justine does not independently file court documents, communicate as the law firm, accept settlements, or execute litigation strategy.

These actions require explicit attorney action.

One of our deliberate design decisions was not to ship autonomous filing. The capability may be technically straightforward, but it conflicts with our position that the attorney must remain the decision-maker.

At the architectural level: every output is attorney-attested work product, and the mandatory review-before-final gate is non-dismissible.

B. Citation accuracy and verification

Under § 6068.1 and CCP § 128.7(b)(2), AI-generated legal citations require verification.

JustineAI's demand-letter workflow employs four passes: draft generation, medical-causation analysis, damages quantification, and four-layer quality assurance.

The fourth QA layer includes citation verification.

Legal citations are checked against the live CourtListener corpus for case name, legal citation, year, and court.

If a citation cannot be resolved, it is removed from the draft and flagged for the attorney of record.

Unverified suggestions are explicitly identified as research leads requiring attorney confirmation.

Only public citation strings are transmitted for citation verification — not client matter content.

You can examine the verification process directly on the public citation-verifier page, which includes sample demand-letter excerpts and authorities linked to CourtListener.

Critical distinction:

Automated citation resolution is not the same as the attorney's personal verification.

The signing attorney retains that responsibility. JustineAI is designed to support that review, not replace it.

C. Client confidentiality

JustineAI's published data-use and architecture policies state:

Customer matter information is never used to train or fine-tune models. Production processing uses inference rather than customer-data training. No production file-upload or batch-training path is used for customer data.

Infrastructure: inference operates inside MindHYVE's Azure tenant, using Managed Identity for authentication. No customer content is transmitted through third-party model-vendor SDKs on the production path. Frontier-model providers operate under contractual restrictions concerning inference-content retention and training.

Data governance: US data residency by default. Azure foundational compliance certifications and controls are referenced, including ISO 27001, SOC 1/2/3, and HITRUST.

D. Supervision and audit evidence

JustineAI records AI-initiated actions in a structured audit log, with action, actor, timestamp, and matter reference.

The audit store is deliberately designed not to contain substantive matter content.

Records are retained for the applicable contractual period and described as exportable for litigation discovery, malpractice insurance, and ethics review.

Why does this matter? When a court, insurer, or disciplinary authority asks what supervision actually occurred, the firm should be able to produce records rather than merely state that a review took place.

We believe transparency about limitations is essential.

It does not autonomously file documents. And we do not intend to introduce autonomous filing that bypasses attorney control.

It does not automatically generate the statutory court-disclosure statement. The § 6068.1 disclosure obligation remains the firm's responsibility. We do not currently publish a capability that independently discharges it.

It does not replace personal citation verification. The citation-verification pipeline identifies authorities that resolve and flags unsupported citations. But CCP § 128.7(b)(2) still requires the responsible attorney's personal verification.

Technology can support the attorney's work. It cannot assume the attorney's professional responsibilities.

If your California law firm is reviewing its AI compliance posture before January 1, 2027, we welcome a discussion.

We can walk through JustineAI's attorney-attested architecture and explain how it maps to the new statutory duties — including the requirements that remain the law firm's direct responsibility.

Explore JustineAI™

Personal Injury Edition — justinegrid.ai/pi

Workers' Compensation Edition (California applicant-side practice) — justinegrid.ai/wc

Citation Verification Technology — justinegrid.ai/technology/citation-verifier

AI & Data-Use Policy — justinegrid.ai/trust/ai-data-use

THE AI REASONS. THE ATTORNEY DECIDES.

Informational advisory only. Not legal advice. No attorney-client relationship is created. Attorneys should independently verify the applicable law, professional rules, and guidance before relying on this material.

This essay is general commentary, not legal advice.

First published on the MindHYVE Insights Substack on October 18, 2026. Lightly edited for house style. Read the original.