ABA Formal Opinion 512, Explained in Plain English

What Opinion 512 is, and what it is not

On July 29, 2024, the ABA Standing Committee on Ethics and Professional Responsibility released Formal Opinion 512, its first opinion on generative artificial intelligence. Nearly two years later it remains the controlling national reference: every subsequent ABA formal opinion has addressed other topics, and the ABA’s AI work since has come through its Task Force on Law and AI (whose final report landed in December 2025) rather than through new ethics opinions.

Two things Opinion 512 is not. It is not binding law; ABA opinions interpret the Model Rules, and only your state’s rules and disciplinary authority bind you. And it is not an AI-specific rulebook; the committee deliberately concluded that the existing Model Rules already cover AI use. That framing has held up: of the state instruments issued since, nearly all apply existing rules rather than writing new ones. California’s proposed AI-specific rule amendments, still pending as of June 2026, would be the first true exception.

The six duties

Competence (Rule 1.1). You do not need to become a machine-learning engineer, but you must have a reasonable understanding of the capabilities and limitations of the specific tools you use, on your own or through colleagues and experts. Because the technology moves, this is a continuing duty, not a one-time checkbox. Several states have pushed the same point further: Kentucky’s E-457 notes that declining to use available AI tools may eventually implicate competence in the other direction.

Confidentiality (Rule 1.6). Before putting any information relating to a representation into an AI tool, you must evaluate the risk that the information will be disclosed to or accessed by others, which means actually reading the tool’s terms: does it train on inputs, who can see prompts, how long is data retained. Opinion 512 says lawyers should obtain a client’s informed consent before inputting information relating to the representation into a tool that may expose it. State instruments are split on how far that goes; see the confidentiality guide for the spectrum from anonymize-and-proceed to West Virginia’s written-consent requirement.

Communication (Rule 1.4). Whether you must tell a client you are using AI depends on the circumstances: the sensitivity of the information, the terms of the engagement, and whether the client would reasonably want to know. Routine, embedded AI (the kind inside mainstream research platforms) generally does not trigger a disclosure duty; feeding a client’s transaction documents into a chatbot is a different conversation.

Candor and meritorious claims (Rules 3.1, 3.3, 8.4). The opinion lands where every court that has sanctioned AI-fabricated citations landed: the signing lawyer owns the filing. AI output that goes to a tribunal must be verified against primary sources first. The sanctions caselaw that prompted this, starting with Mata v. Avianca, is covered in the candor and sanctions guide.

Supervision (Rules 5.1 and 5.3). Managerial lawyers must establish clear policies on the firm’s permitted AI use, and supervising lawyers are responsible for AI use by subordinate lawyers and nonlawyer staff, the same way they are responsible for a paralegal’s work. This is the duty that makes a written firm AI policy effectively unavoidable; the policy requirements guide maps what that policy must contain.

Fees (Rule 1.5). The cleanest line in the opinion: bill the time you actually spend (prompting, reviewing, correcting), do not bill the time the tool saved, and do not charge clients for your own learning curve absent a specific agreement. AI-related costs can be passed through only on the same reasonableness-and-disclosure terms as any other expense. Virginia’s LEO 1901 later added a notable wrinkle for alternative fee arrangements; see the billing guide.

How the states have built on it

Opinion 512 arrived after several states had already spoken (California’s November 2023 guidance came first; Florida’s Opinion 24-1 followed in January 2024) and has since become the anchor citation for the instruments that came after. Mississippi’s Opinion 267 adopted its trust-but-verify analysis outright. Texas Opinion 705 tracks the same four-duty core. Oregon’s 2025-205 extends the confidentiality analysis with an open-versus-closed-model distinction that is more operational than 512’s framing. The full set, honestly labeled by type and linked to primary sources, lives on the tracker.

A 512 compliance checklist

A firm that can answer yes to six questions is substantially aligned with the opinion. Is every AI tool in use on a vetted, approved list with its data-handling terms documented? Does anyone entering client information into a tool know what that tool does with inputs, and has the client consented where required? Is every output reviewed by a human with knowledge of the matter? Is every citation independently verified before filing? Do bills reflect actual time only? Is there a written policy, with training, that makes all of this someone’s job? The free policy template implements each item.