Legal AI tools and California SB 574: What the new law requires of lawyers using AI, and how GC AI supports compliance

California's SB 574 is the first state law to set statutory rules for how attorneys and arbitrators use generative AI. Lawyers may not delegate the practice of law to AI. They may not enter confidential or personal information into AI systems that don't restrict access. They must verify AI outputs, personally verify citations in court filings, and disclose AI use to courts. The law regulates lawyers, regardless of the tools they use.

SB 574's confidentiality rule, though, turns a lawyer's choice of AI tool into a compliance question. GC AI is an enterprise platform built for use by and under the direction of attorneys. It restricts access to customer data through binding confidentiality terms, a no-training commitment, and organization-level data isolation. These protections map directly to what SB 574 requires.

Lawyers are adopting AI at scale, and regulators are catching up. California law has long required lawyers to protect client confidences, and before now, interpretation of how that duty applies to AI has come in the form of ethics opinions and bar guidance. On September 30, 2026, Governor Newsom signed SB 574 (Umberg), which writes AI-specific rules into statute for the first time. The law takes effect on January 1, 2027.

In-house teams should pay attention for two reasons. First, the law reaches every California attorney who uses AI, including lawyers who never appear in court. Second, the application of SB 574's confidentiality rule depends on the terms and controls of the AI tool. The duty to protect client information was already law. SB 574 now says what that duty requires of an AI tool. Before, vendor diligence rested on ethics opinions. Now it is how a lawyer shows compliance with a statutory standard.

What SB 574 requires of attorneys

SB 574 puts its rules in two places. First, the core duties for every attorney go into the State Bar Act, the law that governs lawyer licensing and discipline. Second, the citation rule goes into Code of Civil Procedure Section 128.7, California's version of Federal Rule 11. A violation there can lead to court sanctions. A separate rule covers arbitrators.

1. No delegation. “An attorney shall not delegate the practice of law to generative artificial intelligence.” The law doesn't say what counts as “delegating,” or what counts as “practicing law.” It does say lawyers keep their existing duty to be competent and diligent. The practical reading: AI can help with research, drafting, and review, but the lawyer makes the legal calls.

2. Confidentiality. An attorney who uses AI may not enter confidential, personal identifying, or other nonpublic information into a system unless access is “restricted to the attorney and persons authorized by the attorney under obligations to protect the confidentiality of the information.” “Personal identifying information” is defined broadly. It includes dates of birth, medical or psychiatric information, financial information, account numbers, and the addresses and phone numbers of parties and witnesses. This builds on the existing statutory duty in Business and Professions Code Section 6068(e) to keep client confidences. What's new is a concrete test for AI tools. The question is no longer only whether the lawyer protected the client's secrets. It is also whether the tool's access controls meet the statute's standard.

3. Verification of outputs. Attorneys must take reasonable steps to verify the accuracy of AI outputs, “including, but not limited to, the accuracy of all case and statutory citations.” They must also correct any erroneous or hallucinated output in material they use.

4. Personal verification of court citations. A filing in any California court may not contain citations that “an attorney responsible for submitting the pleading has not personally verified, including any citation provided by generative artificial intelligence.” The rule covers every citation, not only AI-generated ones. Because it sits in Section 128.7, violations carry that section's sanctions. Absent exceptional circumstances, a law firm is jointly responsible for violations by its partners, associates, and employees. Section 128.7 does not apply to disclosures and discovery requests, responses, objections, and motions.

5. Disclosure. Attorneys must disclose AI use to the court “for all documents submitted to the court.” They must also consider whether to disclose AI use for content provided to the public.

Arbitrators get their own rule

Under SB 574, an arbitrator may not hand any part of a decision to a generative AI tool. An arbitrator who wants to rely on AI-generated information from outside the record must tell the parties first, and, where practical, permit the parties an opportunity to respond. That's worth raising when you pick arbitrators or draft arbitration clauses.

SB 574 builds on existing guidance

The duties in SB 574 are familiar. The law codifies California Judicial Council guidance on verification, correcting hallucinations, and disclosure. Bar guidance has long applied lawyers' existing confidentiality duty to AI. ABA Formal Opinion 512 says lawyers must evaluate confidentiality risks before entering client information into an AI tool, and must review its terms of use and data-handling policies. State bars, including California's, have advised that lawyers must not enter confidential information into AI tools that lack adequate confidentiality protections.

Courts already sanction lawyers for fake citations. What SB 574 changes is where the duty comes from and how far it reaches. The citation rule is now written into Section 128.7, the statute courts use to sanction improper filings. That section already holds law firms jointly responsible for their lawyers' violations, and that responsibility now reaches unverified citations.

Consumer-grade AI tools are the likeliest gap. Free and standard consumer tiers of products like ChatGPT and Claude generally lack the confidentiality safeguards of corporate-grade and legal-specific platforms. They often use user data for model training unless the user opts out. A tool that trains on inputs or reserves broad rights to share them is hard to square with SB 574's access-restriction requirement.

Gaps in the statute

SB 574's text leaves several questions open. Until the State Bar, the courts, or the Legislature answers them, lawyers will have to make their own judgment calls:

• What counts as “delegation.” The statute doesn't define it.
• Which attorney must verify. The statute says “an attorney responsible for submitting” a filing. That suggests one attorney may be enough. It's silent on filings with several signers.
• A cross-reference that may have shifted. Section 128.7 has long barred monetary sanctions against a represented party for frivolous legal arguments. That shield points to “paragraph (2).” SB 574 renumbered the section and put the new citation rule in paragraph (2). Read literally, the shield now protects clients from fines for unverified citations but no longer for frivolous legal arguments. Courts may read it to preserve the original intent. Until they do, represented parties may face more exposure.
• What a disclosure must say. The statute doesn't set a format or say how much AI use triggers it.

How GC AI supports attorneys under SB 574

GC AI is built for legal professionals and for use by and under the direction of attorneys. Here is how its design maps to each attorney duty.

1. Confidentiality: access restricted to the attorney and authorized persons. SB 574 asks whether access to what the attorney enters is restricted to the attorney and authorized persons bound by confidentiality. GC AI's Services Agreement is built around that restriction:

• Customer data is Confidential Information. All nonpublic information shared with GC AI, including data, documents, and queries, is “Confidential Information.” GC AI may not disclose it for any purpose other than performing the Services. GC AI must “take all necessary and reasonable precautions to prevent the disclosure of Confidential Information to any unauthorized third parties.” (Sections 5.1, 5.2)
• No model training. GC AI will not use Customer Data to train or fine-tune generative AI models. (Section 4.9) GC AI also maintains zero-data-retention agreements with its LLM providers wherever feasible, as described in its security FAQs.
• Segregated by organization. Customer Content is strictly segregated. Users outside the customer's organization cannot access it. See the security FAQs. Within an organization, chats are private to their creator unless the user shares them.
• Authorized users only. Only registered users within the customer's organization may access GC AI, and the customer controls who those users are. (Section 2.1) That matches the statute's “persons authorized by the attorney.”
• Compelled disclosure protections. If GC AI is ever legally compelled to disclose Confidential Information, it must promptly notify the customer before disclosure if time allows, unless the law prohibits notice. It must also help the customer seek a protective order and disclose only the minimum required. (Section 5.3)
• Technical security. Data is encrypted with AES-256 at rest and TLS 1.2+ in transit. GC AI is SOC 2 Type II certified.

The statute's broad definition of personal identifying information matters here. In-house lawyers handle financial data, employee information, and witness details every day. Review both confidentiality terms and restrictions on permitted inputs before entering that information. GC AI's Services Agreement, Section 3.3.9, prohibits personally identifiable health data and certain other restricted data.

2. Verification: citations you can check. SB 574 requires reasonable steps to verify outputs, including every case and statutory citation. GC AI's research answers link citations to their sources, including official statute and regulation text. Case law answers come from a database of more than 14 million US court opinions, and lawyers can open the full text of each one. Citations to uploaded documents point to the exact passage. These links make verification faster. They don't replace it. For court filings, the responsible attorney still has to read and confirm each authority personally.

3. No delegation: the lawyer decides. GC AI drafts, researches, and analyzes. The lawyer reviews the work and decides what to adopt. Lawyers using GC AI are developing and giving legal advice. The tool is not giving the advice in their place. Playbooks let the legal team set the positions the AI applies, so the team's judgment drives the review.

4. Disclosure: the attorney's call. GC AI does not file anything or make court disclosures. Each attorney decides how to disclose AI use for each filing, following the court's requirements.

Before January 1: Three questions for every AI tool your team uses

SB 574 makes picking a tool part of compliance. Before the law takes effect, ask these three questions about every AI tool your lawyers use, including tools they found on their own:

• Who can see what we enter? The statute requires access limited to the attorney and people the attorney authorizes, all bound to confidentiality. Look for that in the contract, not the marketing page.
• Does the vendor train on our data? A tool that learns from your inputs is hard to square with an access restriction.
• Can we check its citations quickly? Verification is the lawyer's job. Citations linked to their sources make that job workable at volume.

GC AI was built to answer all three: contractual confidentiality, no training on customer content, and source-linked citations. The verification and disclosure duties stay with the attorney, and no vendor can take them on. Choose your tools accordingly. Your AI doesn't have a bar number.

This post reflects GC AI's reading of SB 574 as signed and is general information, not legal advice.

Learn more at the GC AI Security FAQs and Trust Center.

---