ABA Formal Opinion 512 vs. State Bar AI Rules

ABA Formal Opinion 512 issued in July 2024. It is the floor for attorney AI ethics in the United States. Six existing Model Rules duties get restated as applied to generative AI use. No new rules emerge. Existing Rule duties already required everything the opinion now spells out.

State bars that have ruled on the question after July 2024 mostly track ABA 512. A few have gone further. Where states diverge is predictable: disclosure to courts, billing for AI-assisted work, confidentiality of client data, supervision of AI vendors. This post reads the where-they-went-further axis. The structure follows ABA 512’s six duties.

A companion piece, the ABA Opinion 512 compliance guide, walks through what each rule says and includes a 12-item compliance checklist. This post assumes you know the rules in outline and want to know how they bind in your state. Multistate firms should read both.

Quick orientation: ABA 512’s six duties

Six Model Rules get restated. Under Rule 1.1, competence requires understanding the AI tools you use, including their limits and hallucination risk. Rule 1.6 confidentiality bars entering client information into AI tools that may train on or expose that data. Material AI use triggers a client-communication duty under Rule 1.4, with a fact-specific threshold. Billing an hour for ten minutes of AI-assisted work violates Rule 1.5 on reasonable fees. Partners and supervising lawyers carry Rules 5.1 and 5.3 supervision duties for subordinate and non-lawyer AI use. Verifying AI output is the lawyer’s duty under Rule 3.3 candor.

Disclosure to courts is not required by ABA 512 as a default rule. That question is left to local court rules.

State-by-state comparison: the bar opinions

The table below lists state bar AI ethics opinions in approximate effective-date order; each state’s page on the state tracker carries its current status. California is included for context even though its Practical Guidance is technically informal: in practice, California attorneys treat it as the operative standard, and the State Bar of California has proposed amendments to the Comments to Rules 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3 that would carry parts of it into the rules’ official commentary. Under California Rule 1.0(c), Comments “are not a basis for imposing discipline” but guide interpretation of the rules. A second public comment period on that package closed 2026-08-06, with no Board of Trustees action published as of 2026-09-28.

Other states have informal guidance, working groups, or court-rule overlays that practicing attorneys must factor in. Examples: Illinois, New Jersey, and Michigan. The full 50-state-plus-DC matrix lives on the tracker index.

Where states went further than ABA 512

Disclosure and informed consent

ABA 512 leaves disclosure to local court rules. Florida’s Opinion 24-1 is often read as going further, but its consent language is a recommendation tied to confidentiality (Rule 4-1.6), not a requirement: “it is recommended that a lawyer obtain the affected client’s informed consent prior to utilizing a third-party generative AI program if the utilization would involve the disclosure of any confidential information.”

New York City Bar Op 2024-5 is similar in effect: it advises informing the client when AI is material but does not impose an informed-consent requirement.

In court, the picture is more nuanced than a simple disclosure rule. New York’s 22 NYCRR Part 161 took effect June 1, 2026. It sets a system-wide no-disclosure default. Lawyers and parties using AI in court papers are not required to disclose that use. Part 161 then authorizes individual courts to adopt an Appendix A Model Rule. Under the Model Rule, a lawyer who used AI must review the paper for fabricated content, and signing it certifies that review. The certification regime is per-court, not statewide. A firm filing in NY needs to know which courts have adopted the Model Rule. Texas shows how far court rules range within one state. In the Northern District, a brief prepared using generative AI must disclose that fact on its first page (N.D. Tex. Local Civil Rule 7.2(f); tracker entry). The Southern District cautions filers to check AI-drafted papers (S.D. Tex. General Order 2025-04; tracker entry). It holds signers responsible under Rule 11. Some Texas state district courts go further. A July 2026 joint standing order of the 30th, 78th and 89th District Courts of Wichita County (tracker entry) requires a signed certification, attached to each pleading generated using AI, that AI-contributed content was verified through traditional legal sources. It also prohibits AI use in the courtroom and court offices. Texas Rule of Civil Procedure 3a(c) provides: “To be effective, local rules, forms, and standing orders must be published on the Office of Court Administration’s website.” As of September 30, 2026, this order did not appear on that site; the tracker entry records the posting facts. The court orders tracker lists these orders by judge and court. ABA 512 says nothing about these court rules. State bar opinions do not preempt them.

Billing and reasonable fees

ABA 512 restates Rule 1.5. A lawyer cannot bill the client an hour for ten minutes of AI-assisted work plus pretend it was an hour of human work. California’s Practical Guidance is more granular. Two specific instructions appear. First, do not charge for tool subscription fees as if those fees were billable time. Second, bill AI-generated work product at a rate reflecting the actual time spent on review and editing. The human-equivalent rate the document might command if drafted from scratch is not the right measure.

This is a real divergence. Imagine a California firm that bills four hours for AI-assisted research that took the supervising attorney 30 minutes to verify. That firm is exposed under California’s guidance in a way ABA 512 does not directly address. The same firm in Texas, where Op 705 tracks ABA 512 closely on billing, has more latitude in framing.

Confidentiality and vendor diligence

Texas Op 705 has the most detailed factor test for confidentiality. Specific questions are enumerated. Does the tool retain inputs for training? Are outputs accessible to other users? Do the vendor’s terms of service permit secondary use? Is a non-disclosure or business-associate agreement in place? Op 705 also gives examples of tools that pass and fail the test. Most state opinions do not.

NYC Bar Op 2024-5 is comparable in detail. It adds a vendor-supervision frame: AI vendors are treated under Model Rule 5.3, the rule governing supervision of non-lawyer assistants. New Jersey’s Supreme Court Notice (Jan 2024) makes the 5.3 framing explicit. It ties supervision to a duty to verify the vendor’s data-handling policies.

The practical effect: a lawyer in Texas, New York, or New Jersey who pastes a client’s contract into ChatGPT without enabling enterprise privacy controls has a documented bar-rule problem, not just a best-practice problem.

Candor and verification

Verification is the least contested duty. The candor and signature duties that require a lawyer to check cited authority predate generative AI, and the opinions in the table that address candor and verification apply them to AI output. Virginia’s LEO 1901 is the exception: it addresses fees only. The variation that matters is on the consequence side, which lives in court orders and case law rather than bar opinions. The sanctions patterns post tracks the consequence layer.

Fee structure (Virginia’s different view)

Virginia LEO 1901, approved by the Supreme Court of Virginia on November 24, 2025, takes a different view of non-hourly fees than ABA 512. ABA 512 says that if a GAI tool lets a lawyer complete tasks much more quickly, “it may be unreasonable under Rule 1.5 for the lawyer to charge the same flat fee when using the GAI tool as when not using it.” LEO 1901 agrees that time-based billing “can only be based on the actual time spent on a task,” but concludes that “[i]t is not per se unreasonable for a lawyer to charge the same non-hourly fee for work done with the assistance of AI as work done without the use of AI.” Where AI substantially reduces the time spent, it adds that “the client may need additional explanation” of the basis for the fee. The opinion does not cite ABA 512, and it does not address verification duties or malpractice.

For a multistate firm, the practical difference is in flat and other non-hourly fees: a fee that LEO 1901 treats as not per se unreasonable could draw the scrutiny ABA 512 describes in a state that follows ABA 512’s reasoning. Firms using one fee structure across offices may wish to check each state’s guidance.

Supervision

Oregon Op 2025-205 treats firmwide AI use as a supervisory matter under Rule 5.1, with examples of what monitoring looks like in practice. NYC Bar Op 2024-5 and New Jersey Supreme Court Notice (Jan 2024) reach a comparable place via Rule 5.3, framing AI vendors as non-lawyer assistants that supervisory lawyers are responsible for vetting. ABA 512 covers the same territory but at a higher level of generality.

Where ABA 512 is the ceiling and not the floor

A handful of states have gone short of ABA 512, mostly by silence. Illinois has no formal bar ethics opinion on AI; the Illinois State Bar Association’s opinion index lists none through September 2026. Instead, the Illinois Supreme Court announced a Policy on Artificial Intelligence on December 18, 2024, effective January 1, 2025. The Policy is not limited to court personnel. It states that AI use “by litigants, attorneys, judges, judicial clerks, research attorneys, and court staff providing similar support may be expected, should not be discouraged, and is authorized provided it complies with legal and ethical standards,” that “Disclosure of AI use should not be required in a pleading,” and that “All users must thoroughly review AI-generated content before submitting it in any court proceeding” to ensure accuracy. The ARDC followed in October 2025 with a nonbinding implementation guide for attorneys. Because the Policy says the Rules of Professional Conduct “apply fully to the use of AI technologies,” Illinois practitioners work under the existing Illinois Rules, whose Rule 1.1 Comment [8] already covers “the benefits and risks associated with relevant technology,” rather than under an AI-specific rule.

States with no formal opinion include Wyoming, North Dakota, South Dakota, Vermont, and most other small-bar states. Where their courts have spoken, they have mostly restated existing duties rather than added new ones. The District of Wyoming’s General Order 25-01 (May 2025) reminds all litigants of their Rule 11 obligations and states that filing hallucinated authority “is strictly prohibited and subject to sanction.” The North Dakota Court System’s June 1, 2026 public notice permits AI use and does not require disclosure “solely because AI was used.” The trend is toward formal opinions, so a quiet state today may not stay quiet.

Multistate practice: cross-state risk

Take a firm with offices in California, New York, and Texas. Three distinct regimes apply:

• California’s Practical Guidance on billing
• New York’s NYC Bar Op 2024-5 confidentiality emphasis plus Part 161’s per-court certification overlay
• Texas’s Op 705 vendor-diligence factor test

The firm’s AI policy must satisfy the most restrictive rule across the three. In practice that means three things. Billing aligned to actual time spent (the California floor). Vendor diligence with documented contract review (the Texas floor). A per-court check at intake to identify whether each filing court in NY has adopted the Part 161 Appendix A Model Rule (which requires a certification, not a disclosure). A policy that satisfies any one of these states alone leaves gaps in the others.

For multistate firms, the Carrier-Renewal Packet, now in development, is designed to annex the policy to the guidance that applies in each state where the firm practices. For firms with operations in three or more states, consulting runs the gap analysis state by state.

What changes between now and 2026 year-end

Three live items to watch:

1. California Comment amendments. COPRAC’s proposed amendments to the Comments to Rules 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3 would restate the Practical Guidance’s core duties in the rules’ commentary; the rule text itself would not change. The second public comment period closed 2026-08-06, and COPRAC’s September 11, 2026 agenda listed the proposals for possible recirculation or submission to the Board of Trustees. Adoption still needs Board of Trustees action and California Supreme Court approval, and none had been published as of 2026-09-28. Proposed amendment text and current status are on the California state page.
2. Court rules. Individual judges keep issuing standing orders, and in 2026 several state high courts wrote the verification duty into their rules of procedure. Florida amended Rule of General Practice and Judicial Administration 2.515 so that a signer represents that “the legal authorities identified exist and are accurately cited” (tracker entry). Connecticut’s amended Rules of Appellate Procedure now list failure to verify citations produced by generative AI as a ground for sanctions (tracker entry). Indiana’s amended Trial Rule 11 has an attorney’s signature certify “that any cited legal authority is authentic” (tracker entry). The court orders tracker follows these rules alongside judges’ standing orders. Compliance with court rules is a separate sanction risk on top of bar duties.
3. State rule-comment packages. Some states are moving AI language into their rules or commentary rather than issuing opinions. Rhode Island’s Supreme Court acted by order in June 2026. It amended Comment [6] to Rule 1.1 to cover “the benefits and risks associated with existing and developing technology.” It also adopted interim generative AI guidelines for lawyers, which it calls “advisory in nature” (tracker entry). Two other states have packages still in progress. Maryland’s Rules Committee approved AI-related amendments to its versions of Rules 1.1, 1.6, 3.3, and 5.3 in June 2026, including a comment that a lawyer relying on AI output “must independently review, verify, and exercise professional judgment” about it (June 25, 2026 minutes; Maryland page). In July 2026 the Washington State Bar’s Board of Governors agreed to co-propose AI comments to the Rules of Professional Conduct with the Washington Supreme Court Clerk (July 2026 Board recap; Washington page). Both still need their Supreme Courts’ action. Our state tracker updates as they move.

A note on state divergence and bar enforcement

Bar enforcement of AI-related rules is uneven. Cases coded as bar discipline (rather than court sanctions) are a small subset of the sanctions docket. They come from states with and without a formal AI opinion. A bar complaint rests on the Rules of Professional Conduct themselves; an AI opinion explains how those rules apply.

Two factors matter when a firm assesses its risk. First, whether the firm’s home state has a formal opinion that names AI duties specifically. Second, whether the firm files in any court whose standing orders impose disclosure or verification on top of the bar rule. Both layers carry sanctions exposure. A clean record under the bar rule is not a defense if the court rule is what was violated.

This article is for informational purposes only and is not legal advice. State bar rules and court rules vary; consult a licensed attorney in your state for representation-specific questions. All citations are to primary sources; see the state tracker for the latest opinion text.