California Is Turning AI Ethics Guidance Into Disciplinary Rules

What California actually proposed, and why it is a departure

California did not invent new AI duties. It proposed moving existing ones out of guidance and into the enforceable rules that govern discipline, which changes their weight.

On August 22, 2025, the California Supreme Court sent the State Bar a letter directing it to consider incorporating principles from the Bar's 2023 practical AI guidance into the Rules of Professional Conduct ^[1]. COPRAC responded by drafting language and, at its March 13, 2026 meeting, approving proposed amendments to six existing rules for public comment ^[1]. The comment period ran 45 days and closed on May 4, 2026 ^[1]. As of this writing the amendments are proposed and in the review process, not adopted, so nothing described here is yet an enforceable rule.

The design choice is worth noting. Rather than draft a single standalone AI rule, COPRAC wove new language into six rules the profession already recognizes, reflecting a view that AI sharpens existing ethical duties rather than inventing new ones ^[1]. That approach keeps the familiar structure of the rules intact while making clear that competence, confidentiality, communication, candor, and supervision all apply when a lawyer uses generative AI.

The departure is not the subject matter, which is broadly consistent with earlier guidance and with the American Bar Association's national opinion. The departure is the legal status. Advisory guidance describes duties and persuades. A Rule of Professional Conduct binds, and a violation can support discipline. Moving these expectations into the rules changes them from something a lawyer should heed into something a lawyer must follow, with a disciplinary consequence attached.

Advisory guidance versus an enforceable rule: the distinction that matters

Most of the AI direction lawyers have received to date is advisory. Surveys that track this, such as Justia's 50-state review, count more than two dozen jurisdictions that have issued some form of AI guidance for lawyers, most of it advisory rather than binding ^[4]. Advisory guidance is useful and often persuasive, but it does not on its own create a new disciplinable rule. It restates duties the rules already impose and applies them to a new technology.

An enforceable rule is different in kind. When an expectation lives in the Rules of Professional Conduct, a lawyer who fails to meet it can face discipline for violating the rule itself. California's 2023 practical guidance was advisory ^[2]. The 2026 proposal would move the substance of that guidance into the rules ^[1]. The table below lays out the contrast, using California's own documents as the example.

The proposed amendments remain proposed. Describing them as enforceable describes what they would be if adopted, not their status today ^[1].

The six Rules of Professional Conduct the amendments would touch

The proposal reaches six rules, each mapping to a duty the profession already recognizes. COPRAC approved amendments to Rules 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3 ^[1]. Reading them together shows the shape of the intended obligations: verify AI output, communicate with clients, protect confidential information, be candid with courts, and supervise both people and tools. The table summarizes what each rule addresses in the AI context.

The subjects will look familiar to anyone who has read the American Bar Association's national opinion, which reached similar ground through the same duties ^[3]. The difference is that California proposes to write the AI expectations into the binding text of the rules rather than describe them in a formal opinion.

Rule subjects and numbers are drawn from the State Bar's public comment page ^[1]. The rules are proposed amendments, not adopted.

Rule 1.1 and independent verification: competence becomes a checkable step

The competence proposal is the one that has drawn the most attention. As reported, it would require that a lawyer independently review, verify, and exercise professional judgment over any output generated by technology used in representing a client ^[1]. In plain terms, a lawyer cannot treat AI output as finished work. The output has to be checked by a person exercising legal judgment.

This is consistent with the national position. The American Bar Association's Formal Opinion 512 confirms that the duty of competence applies to generative AI and that a lawyer's professional judgment cannot be delegated to a tool ^[3]. What California proposes to add is the enforceable framing: the verification step would sit inside Rule 1.1 itself, so failing to perform it is not just a departure from good practice but a potential rule violation ^[1].

The practical consequence for firms is that verification stops being a private habit and becomes something a lawyer may need to demonstrate. It is one thing to believe an associate checked the AI output. It is another to be able to show that the check happened on a specific matter. A rule that requires independent review invites the question of how a lawyer would establish that the review took place, which is where records become relevant. For a related discussion of citation accuracy, see our piece on fabricated citations in legal filings.

Rule 1.6 and confidentiality: not the same thing as privilege

Rule 1.6 confidentiality is a broad ethical duty. The attorney-client privilege is a narrower evidentiary protection. The AI proposal addresses confidentiality, and the two should not be treated as interchangeable.

The confidentiality proposal touches Rule 1.6, which governs a lawyer's duty to protect confidential client information ^[1]. In the AI context, the concern is concrete: client information entered into a third-party AI system may be stored, processed, or used to train a model unless the arrangement prevents it. The 2023 practical guidance already warned that a lawyer should not input confidential information into a generative AI tool unless the provider will not share or train on it, and that inputs should be handled so they do not identify the client ^[2]. The proposal would move that concern into the enforceable rule ^[1].

It is important to be precise here, because two ideas often get blurred. The duty of confidentiality under Rule 1.6 is broad. It covers information relating to the representation, regardless of source, and it is an ethical duty owed by the lawyer. The attorney-client privilege is a narrower evidentiary protection that shields certain communications from compelled disclosure and that can be waived. They are related but distinct. A confidentiality obligation under Rule 1.6 can apply to material that would never be privileged, and a confidentiality breach is not the same event as a privilege waiver. Firms evaluating AI tools should keep the two concepts separate, because the proposed rule is about confidentiality, not about privilege.

The operational takeaway is that a firm should be able to show what happened to client material when an AI tool was involved. We cover the mechanics of keeping client data isolated in our guide to client confidentiality and AI.

Rule 3.3 and candor: the hallucination problem written into the duty

The candor proposal amends Rule 3.3, which governs a lawyer's duty of candor toward a tribunal, to address the widely reported problem of AI-generated citations that do not exist or do not say what a filing claims ^[1]. The proposal reinforces that a lawyer is responsible for ensuring AI-assisted citations are accurate and verifiable ^[1]. This is the failure mode that has generated sanctions and judicial attention across the country, and it is the reason candor is on the list.

A careful note on language matters here. No responsible approach claims that generative AI is hallucination-free. Models can and do produce fabricated or misgrounded authority, which is exactly why a verification duty is being written into the rules. The point of Rule 3.3 in this context is not that AI can be made perfectly reliable. It is that the lawyer, not the tool, is answerable for what reaches the court, and that the lawyer must confirm cited authorities are real and accurately represented before filing.

For firms, the response is a verification step that produces a record. A certification that cited authorities exist and are accurately cited addresses the specific risk courts have flagged, and it creates evidence that the check occurred. It does not certify that the legal argument is correct or that every proposition is supported, which remain matters of lawyer judgment. That honest scope is the difference between a claim a record can support and a claim it cannot.

Rules 5.1 and 5.3: supervision and firm-level AI governance

The supervision proposals reach Rule 5.1, on the responsibilities of managerial and supervisory lawyers, and Rule 5.3, on responsibilities regarding nonlawyer assistants ^[1]. As reported, managerial lawyers would be expected to make reasonable efforts to establish procedures governing the use of artificial intelligence in accordance with the Rules of Professional Conduct ^[1]. In other words, AI governance would become a firm-level obligation, not only an individual one.

This mirrors the national position that supervisory duties extend to generative AI, which the American Bar Association confirmed in Formal Opinion 512 ^[3]. The California proposal would give that expectation the force of an enforceable rule at the firm level: a firm would be expected to have procedures, and a managerial lawyer could be answerable for their absence ^[1].

The practical implication is that a written AI policy moves from a nice-to-have toward an expected control, and that the policy needs to be more than a document in a shared drive. A supervisory rule invites the question of whether the firm's procedures actually operated on a given matter. We walk through how to build a policy that a firm can demonstrate it followed in our guide to a defensible law firm AI policy.

Why this matters for firms outside California

A firm in another state might reasonably ask why a California proposal is relevant to it. There are three reasons. First, California is one of the largest legal markets in the country, and when its rules change, other states often study the result. When one of the leading jurisdictions writes AI duties into enforceable rules, it becomes a reference point for regulators elsewhere.

Second, the direction of travel is broad. Surveys that track ethics guidance count more than two dozen jurisdictions that have addressed AI, and roughly half of states have issued something, most of it advisory so far ^[4]. California moving from advice to enforceable rule is the leading edge of a trend, not an isolated event. Firms that operate across state lines are already navigating a patchwork, and that patchwork is hardening in places.

Third, and most practically, enforceable rules turn should into must, and they create discipline exposure that advisory guidance does not. Whatever the outcome in California, the useful posture for any firm is to assume that verifying AI output, protecting confidential information, being candid about AI-assisted citations, and supervising AI use are expectations they may one day have to demonstrate they met. Building the habit of being able to prove those steps is portable across jurisdictions in a way that reacting to each new rule is not. Sound information security underpins all of this, which we address in our overview of law firm cybersecurity.

What "prove it" looks like when the standard is enforceable

The shift from advisory to enforceable changes the question a firm should ask about its AI use. Under guidance, the question is roughly whether the firm is following good practice. Under an enforceable rule, the question becomes whether the firm can show it followed the rule on a specific matter, because a rule violation is what supports discipline. That is a higher bar, and it is a bar about evidence, not only intention.

In practice, being able to prove compliance means keeping records that show the right steps happened. For competence under Rule 1.1, that means evidence that AI output was independently reviewed and that cited authorities were checked. For confidentiality under Rule 1.6, that means being able to show what happened to client material and that it was not exposed to a system that would store or train on it. For candor under Rule 3.3, that means a record that citations were verified before filing. For supervision under Rules 5.1 and 5.3, that means demonstrable procedures rather than an unwritten understanding.

This is the lane RankShield builds in. RankShield is a technology vendor, not a law firm, and it does not provide legal advice. What it provides is the ability to produce verifiable records: attestations that an AI tool was used within defined limits, that cited authorities were checked, and that client material stayed isolated. Those records do not discharge a lawyer's duties, and they do not decide legal questions such as whether privilege attaches. What they add is evidence, so that when a firm is asked whether its AI governance was real, it can show a record rather than restate an intention. For related vetting steps, see our policy guide.

What to do now while the rules are still proposed

Because the amendments are proposed and not adopted, there is no California rule to comply with today beyond the duties that already exist and the advisory guidance already in place ^{[1] [2]}. That makes this a useful window to prepare rather than react. The sequence below is a general illustration for firms that want to be ready, not legal advice, and it should be adapted with qualified counsel for your jurisdiction.

1. Read the actual proposal, not summariesReview the State Bar's public comment page and the proposed rule text directly, so your understanding rests on the primary source rather than secondhand accounts ^[1].
2. Map your current practice to the six rulesCompare what your firm does today against competence, communication, confidentiality, candor, and supervision, and note where you could not currently prove the step occurred.
3. Separate confidentiality from privilege in your analysisTreat Rule 1.6 confidentiality and the attorney-client privilege as distinct questions when evaluating AI tools, because they protect different things and are waived or breached in different ways.
4. Turn verification into a recorded stepMake independent review of AI output, and confirmation that cited authorities are real, a routine that produces a record rather than an unwritten habit.
5. Put firm-level AI procedures in writingBecause Rules 5.1 and 5.3 point toward firm governance, establish written procedures for AI use and assign supervision, so the control exists and can be shown ^[1].
6. Revisit as the process advancesBecause the amendments are still moving through review, track the outcome and update your approach when the status changes, rather than assuming today's proposal is final.

This sequence is informational and not legal advice. The amendments are proposed and were in public comment; confirm the current status and requirements with counsel and at the State Bar's site ^[1].