Mandatory AI Disclosure Upholds Ethical Standards in Law

California’s new rules for lawyer use of generative AI put a basic professional question in sharper focus: When technology helps produce a court filing, who remains accountable for the result?

California’s Senate Bill 574 requires attorneys to disclose their use of generative AI in documents submitted to courts. The law also requires personal verification of citations, protects confidential client information, and bars lawyers from delegating legal judgment to an AI system. Those requirements create work for firms, especially when AI is already embedded in research, drafting, proofreading, document review, and workflow software.

The disclosure requirement still serves an important purpose. Courts, clients, and opposing counsel deserve a clear record of how a legal document was prepared. That transparency supports the same kind of accountability that aviation and medicine built around checklists, handoffs, and incident reporting. A pilot remains responsible for a flight after relying on automated systems. A physician remains responsible for a treatment decision after consulting software or diagnostic tools. A lawyer remains responsible for a filing after using generative AI.

California’s law takes effect Jan. 1, 2027, according to legal briefings on SB 574. It places several duties on attorneys who use generative AI in legal work:

The statute follows a series of cases involving invented authorities and distorted quotations. A California appeals court sanctioned counsel $5,000 in March after fictitious AI-generated citations appeared in a custody dispute. Similar problems have surfaced in courts around the country. Legal AI systems can produce text that sounds authoritative while inventing a case, misstating a holding, or attaching a real quotation to the wrong decision.

The American Bar Association’s Formal Opinion 512 already tells lawyers to consider duties involving competence, confidentiality, communication, supervision, and reasonable fees when they use generative AI. California’s law adds statutory force and a broad disclosure rule. That difference matters. Professional guidance can shape conduct and discipline. A statute gives courts a direct legal standard to apply.

Scott Meyers, chairman of Akerman LLP, described the distinction plainly: “An ABA opinion is guidance, and SB 574 is law.” He also identified the practical problem that firms now face: AI has become part of tools lawyers use every day. A lawyer may use a platform for research, drafting, editing, citation checking, or document management without making a separate decision to open a chatbot.

California Senate Bill 574, effective January 1, 2027, requires lawyers to disclose generative AI use in court filings, personally verify every citation, and sign filings in their own names. The law also protects confidential information and bars lawyers from delegating the practice of law to AI. Although the requirements will create costs and ambiguity, they establish a visible chain of responsibility for legal work produced with rapidly changing tools.

A simple policy can cover a lawyer who asks a public chatbot to draft a motion. The harder cases involve background automation.

A legal research platform may use a generative model to summarize authorities. A word processor may suggest revisions. A contract system may extract provisions or produce a comparison. A document platform may classify files or generate a first draft. A firm approved tool may use a large language model behind the scenes while presenting the result through a familiar interface.

The statute’s language creates pressure to define these activities carefully. Peter Salib, an assistant law professor at the University of Houston Law Center, said the disclosure requirement applies to all documents submitted to court and contains no exception for minor uses. Under that reading, a firm that uses AI to research, draft, or edit a brief needs a standard disclosure and a way to identify which filings the technology touched.

That broad approach has a practical advantage: it gives lawyers a clear reason to record AI involvement. It also creates an administrative burden. A firm could end up tracking a grammar suggestion with the same system used to generate several pages of argument. A narrow threshold would reduce paperwork, but it would require lawyers to decide where the threshold lies and prove that a particular use falls below it.

Daniel W. Linna Jr., director of law and technology initiatives at Northwestern University, has criticized the requirement as burdensome and overbroad. He has also suggested that lawyers could challenge it on First Amendment grounds. That legal debate may continue, but firms still need an operating policy while the rule remains in place.

A disclosure statement does more than satisfy a filing requirement. It gives the court a signal about the level of review that should have followed the use of the tool. It also allows opposing counsel to raise a focused concern if a filing contains a suspicious citation, quotation, or factual assertion.

A short statement might say that generative AI assisted with drafting or editing and that the responsible attorney independently reviewed the filing, verified its citations, and accepts responsibility for its contents. The exact wording will depend on court rules, firm policy, and future guidance. The key point is that the statement should describe the role of the technology without implying that the tool acted as counsel.

Clients may also benefit from the same transparency. The ABA guidance emphasizes communication about significant decisions affecting a client’s matter. If AI changes the cost, speed, staffing, or nature of legal work, the lawyer may need to explain that use to the client. A client could reasonably care whether a document was drafted by an attorney, assembled with an approved tool, or generated through an unrestricted public service.

Trust in professional services depends partly on knowing where judgment entered the process. A client does not need a technical lecture about model architecture. The client needs a useful explanation of what the tool did, what data it received, what review followed, and how the firm protected confidential information.

Mandatory disclosure is justified because speed and plausibility make AI errors unusually difficult to detect before they reach a court. The rule should not turn every AI-assisted filing into a presumption of misconduct, but it should force firms to know which tools they use, protect client information, and build a deliberate human review point before a filing carries a lawyer’s name.

SB 574 requires the responsible attorney to personally verify citations, including citations supplied by AI. Ariel House, a partner at Baker Botts, has noted that this requirement may impose a significant burden. Many firms rely on paralegals to check cases, quotations, exhibits, and cross-references under an attorney’s supervision. The attorney’s personal verification adds a final step that cannot be delegated entirely.

That step changes how firms should allocate time. A lawyer may still assign a preliminary citation check to a paralegal or associate. The responsible attorney must then review the source and confirm that the citation supports the proposition in the filing. The process should cover more than whether a case exists. It should examine:

AI tools can help with parts of this process, but the law places the final responsibility on an attorney. That division resembles safety procedures in other high-stakes fields. Automated alerts can identify a possible problem. A qualified professional must decide what the alert means and whether the decision is safe to carry forward.

Small firms and solo practitioners may feel this requirement most sharply. A large firm can build a review team, buy verification software, and appoint an AI governance group. A solo lawyer may have to handle research, drafting, citation review, filing, and disclosure personally. Compliance policies therefore need to account for staffing realities rather than assume that every practice has an enterprise support structure.

The law’s restriction on entering confidential or nonpublic client information into public AI tools reinforces a long-standing professional duty. A prompt can expose names, medical details, financial records, trade secrets, litigation strategy, or privileged communications. Once information enters a system, the lawyer needs to understand how the provider stores it, uses it, retains it, and permits access to it.

A firm policy should classify information before an employee places it into an AI system. One workable approach could separate information into three levels:

The classification alone does not solve the problem. Lawyers also need training, vendor review, access restrictions, retention rules, and an incident process for accidental disclosures. A tool marketed as secure still requires a careful review of its terms and technical controls.

Client consent may be appropriate in some situations, particularly when AI use materially affects the representation or billing. Consent does not relieve a lawyer of the duty to protect information. It does create a clearer conversation about the client’s expectations and the firm’s methods.

The next challenge may involve systems that take actions rather than produce text. Legal technology companies are developing agentic tools that can search records, organize evidence, prepare drafts, send requests, update matter systems, or carry out multistep workflows. Those tools may save time, but they also create more points where an automated mistake can affect a client or a court filing.

Meyers said legal AI is moving from drafting documents to taking actions on a lawyer’s behalf. Akerman designs its agentic tools so a lawyer reviews the work and makes the final decision. That human checkpoint is important, especially when an action changes a case record, communicates with a third party, or creates a representation to a court.

A useful firm policy should identify actions that require direct attorney approval. Examples could include:

The policy should also preserve an audit trail. If an AI agent searches a database, prepares a draft, or makes a recommendation, the matter file should show what the system did and what the attorney reviewed. A record helps the firm investigate errors and demonstrate compliance if a court asks how a filing was prepared.

Mandatory disclosure may feel uncomfortable because lawyers often treat their methods as private work product. The rule changes that expectation for court filings. It asks attorneys to identify a material part of the production process so the court can evaluate the filing with proper context.

That transparency can encourage better internal habits. When lawyers know they must disclose AI use, they have a reason to identify approved tools, document review steps, and maintain clear responsibility for each filing. The disclosure requirement can also make it harder for a firm to treat AI as an invisible labor substitute while leaving professional accountability undefined.

Aviation offers a useful comparison because its safety systems rely on routine reporting and shared responsibility. Pilots, maintenance crews, and air traffic personnel use checklists and records that expose potential failures before they become accidents. Medicine uses similar safeguards through informed consent, charting, medication checks, and review of adverse events. The legal profession has its own version of this duty: accurate filings, protected client information, documented supervision, and a lawyer who stands behind the result.

The comparison has limits. Court filings involve judgment, advocacy, and constitutional rights rather than mechanical systems or clinical treatment. The underlying lesson still transfers. High-stakes professionals preserve public trust by making responsibility visible and by building review into ordinary work.

A policy written in broad terms will leave lawyers guessing. Firms preparing for the California rule should establish a process that answers operational questions before a filing reaches court.

A useful plan could include the following steps:

A firm may also want a central register for court filings that involved AI. Such a register creates a record without forcing every lawyer to reconstruct the workflow months later. Smaller practices can use a simple matter checklist if a full software system is impractical.

California’s law will test whether legal technology operates under clear professional control. Disclosure alone cannot prevent hallucinations, protect every client secret, or guarantee a sound legal argument. It can create an accountability point that connects the tool, the lawyer, the firm, and the court.

That connection matters as AI becomes more deeply embedded in legal work. Courts need filings they can trust. Clients need to understand how their matters are handled. Lawyers need workflows that preserve independent judgment while allowing useful technology. Mandatory disclosure gives each group a clearer view of the process and reinforces the standard that has always governed legal practice: the attorney who signs the work remains answerable for it.