ABA Formal Opinion 512 (2024): Generative AI Duties for Lawyers

Everyone remembers Mata v. Avianca for the wrong reason.

In June 2023, Judge P. Kevin Castel of the Southern District of New York fined two lawyers $5,000 under Rule 11 for filing a brief built on six cases ChatGPT had invented. The lawyers had pasted fabricated citations into a federal filing and, when challenged, asked the same chatbot whether the cases were real. It said yes. They believed it.

The story went viral because it was embarrassing, and the lesson everyone took home was simple: check your citations.

That lesson is correct, and it is also the smallest part of what happened. When the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512 on generative artificial intelligence tools on July 29, 2024, it did not write a hallucination rule. It wrote a map of six existing duties and explained how each one applies when you hand work to a machine.

The hallucination problem is on that map. So are four other obligations that bite even when your AI never invents a single case, even when nothing gets filed, even when you do everything Mata teaches. Those are the duties most lawyers skip, and they are the ones most likely to cost you a client relationship or a malpractice claim.

Short answer: ABA Formal Opinion 512, issued July 29, 2024 by the Standing Committee on Ethics and Professional Responsibility, creates no new rules. It applies existing Model Rules to generative AI: competence (1.1), confidentiality (1.6), communication (1.4), candor and meritorious claims (3.1, 3.3, 8.4(c)), supervision (5.1, 5.3), and reasonable fees (1.5). You must understand the tool, verify its output yourself, get informed consent before client information goes into a self-learning tool, and bill only the time you actually spent.

TL;DR

• ABA Formal Opinion 512 (July 29, 2024) does not create new rules. It applies existing Model Rules to generative AI: competence (1.1), confidentiality (1.6), communication (1.4), candor and meritorious claims (3.1, 3.3, 8.4(c)), supervision (5.1/5.3), and fees (1.5).
• The opinion refuses to draw bright lines. How much you must verify is "fact-specific" and depends on the tool and the task. There is no ABA checklist that makes you compliant.
• The two duties almost everyone ignores are fees and confidentiality. You generally cannot bill clients for time spent learning a tool, and you need informed consent before feeding client information into a self-learning AI.
• Mata v. Avianca is the origin story, not the whole story. Candor and verification matter, but 512 is broader and quieter than the sanctions headlines suggest.
• Adoption is still early: roughly 30% of lawyers report using AI, and about three in four cite accuracy fears as the top barrier (2024 ABA Legal Technology Survey, released March 2025). The compliance burden has shifted to your tool selection and your workflow.

Part of our legal AI verification and hallucination guide series.

For related verification / hallucination / vendor-trust coverage, see Are ABA Formal Opinions Binding? What Opinion 512 Means for Your AI Workflow and How to Verify AI Legal Citations Before You File (ABA 512 Checklist).

Why 512 Had to Exist

The ABA does not issue formal opinions because a technology is new. It issues them because lawyers are already in trouble and the existing rules need translating.

By mid-2024, Mata was no longer a one-off. The pattern had repeated across jurisdictions: a lawyer uses a general-purpose chatbot for research, the chatbot produces confident, well-formatted, completely fictional case law, and the lawyer files it without looking.

We documented that arc and the verification protocol that prevents it in our piece on AI hallucinations and sanctions, so I will not relitigate the case roll-up here. The point for 512 is narrower: the profession had demonstrated, in open court, that it did not understand what these tools are.

The full text of the opinion lives on americanbar.org (the page blocks automated traffic but is publicly live); a clean walkthrough of its holdings is in the National Conference of Bar Examiners' Bar Examiner coverage, which I leaned on for the rule-by-rule reading below.

That is the framing 512 quietly adopts. Generative AI is not a research database that occasionally errs. It is a delegate. And lawyers have always known how to supervise delegates, because the Model Rules have governed how partners oversee associates and how firms oversee paralegals for decades.

512's move is to say: treat the AI the way you would treat a brilliant, fast, eager first-year associate who has never been to law school, cannot tell you when it is wrong, and will confidently make things up to please you. You would never file that associate's work without reading it. You would never let that associate decide what to tell the client. You own the output.

Hold that analogy. It is the through-line for all six rules.

The five duties Opinion 512 maps onto AI use: competence, confidentiality, communication, candor, and fees.

Rule 1.1 (Competence): You Have to Understand the Tool, Not Just Use It

Competence under 512 has two layers, and most coverage only mentions the first.

The first is the obvious one: you must understand the benefits and risks of the generative AI you use, the same way Comment 8 to Rule 1.1 already requires technological competence. You do not need to be able to build the model. You do need to understand, at a working level, what it does well, where it fails, and why.

The second layer is the one that turns competence into a verification duty. Because the AI can hallucinate, the opinion says lawyers must independently verify its output before relying on it. But, and this is the deliberate part, 512 refuses to say how much verification is enough.

The required level of review is "fact-specific" and "depends on the GAI tool and the specific task." A summarization task on a document you provided is different from open-web case-law generation, which is different again from a tool grounded in a retrieved corpus with linked citations.

This is where the ethics opinion stops being a checklist and becomes a tool-selection problem. The ABA will not tell you a number. So the burden lands on the workflow you choose.

A tool that retrieves from real opinions and shows you the source it relied on (retrieval-augmented generation, not a model reciting from training data) makes your verification cheap and fast: open the citation, read the passage, confirm it says what the tool claims. A tool that generates a citation from nothing forces you to do the entire research task again by hand, at which point you have to ask what you bought.

The unbounded verification duty in 512 is exactly why grounded-research workflows are the cheapest defensive choice: the cheaper you make verification, the more defensible your practice becomes. (How RAG changes the failure modes is its own topic, covered in how AI legal research works.)

Rule 1.6 (Confidentiality): The Sleeper Duty Most Lawyers Get Wrong

Here is the first rule that bites even when nothing hallucinates and nothing gets filed.

512 addresses what happens when you type client information into a generative AI tool. The headline holding: before you input information relating to a representation into a self-learning GAI tool, one that uses your inputs to train itself or improve the model for other users, you generally need the client's informed consent.

Read that twice, because two words do a lot of work.

"Self-learning" matters because the risk is that your privileged facts become training data, surfacing later in some other user's session or persisting on a vendor's servers in ways you cannot control. A tool with a no-training, zero-retention posture is a different confidentiality animal than a consumer chatbot whose terms reserve the right to train on your inputs.

"Informed consent" matters because the opinion is explicit that boilerplate engagement-letter language is not enough. A generic "we may use technology including artificial intelligence" clause does not constitute informed consent for feeding a specific client's specific confidences into a specific self-learning system. Informed consent means the client understands the actual risk to their actual information.

This is the rule lawyers skip because it has no dramatic court scene attached. Nobody gets sanctioned in open court for it. But it is the one most likely to blow up a client relationship, because the moment a sophisticated client learns their deal terms went into a consumer chatbot's training pipeline, you have a trust problem that no apology fixes.

The practical defense is to know exactly where your inputs go before you type them. Read the vendor's terms on training, retention, and sub-processors; prefer tools whose data posture you can actually point a client to. The longer treatment of vendor posture lives in where your legal AI data actually goes.

Rule 1.4 (Communication): When You Have to Tell the Client

Communication is the quiet companion to confidentiality, and 512 handles it with characteristic restraint.

The opinion does not impose a blanket duty to disclose every use of AI to every client. Using a tool to clean up grammar or organize a document is not, by itself, something you must announce. But disclosure becomes necessary when the client would reasonably want to know: when the AI's use is material to the representation, when the client has asked about your methods, or, critically, when you need their informed consent under Rule 1.6 anyway.

The practical read: communication and confidentiality move together. If you are about to put client information into a self-learning tool, the consent conversation under 1.6 is also your communication moment under 1.4. Handle them as one.

Where AI use is incidental and the data is contained, you generally do not owe a disclosure, but you do owe yourself an honest assessment of which bucket you are in.

Try Vaquill AI on your matterBook a demo

Rule 3.3 (Candor) and Rules 5.1/5.3 (Supervision): The Mata Duties

These are the rules everyone already associates with AI, so I will be brief and point you to the deeper treatment.

Candor to the tribunal (3.3) is the Mata rule. You cannot present fabricated authority to a court, and your ignorance of the fabrication is not a defense if you failed to verify. This is the duty Schwartz and LoDuca violated, and it is the one that produces sanctions. 512 ties this section to two neighbors: Rule 3.1 (do not advance frivolous claims or contentions, which AI-invented authority can do on its own) and Rule 8.4(c) (do not engage in conduct involving misrepresentation). Read together, they make the verification duty explicit at the courthouse door, not just the research desk.

The verification protocol that actually satisfies it (what to check, in what order, with what evidence) lives in our sanctions deep-dive, not here.

Supervision (5.1 and 5.3) is the structural rule, and it is more interesting than it looks. 512 treats GAI use as a managerial responsibility. Partners and managers must establish policies that give reasonable assurance the firm's AI use conforms to the rules.

Rule 5.3 historically governs nonlawyer assistants, and the opinion analogizes the AI tool to exactly that: a nonlawyer assistant whose work you are responsible for. This is the brilliant-first-year-associate analogy made into a rule.

You cannot delegate your professional judgment to the tool, and you cannot let your associates do it either. The firm needs a written AI policy, training, and a verification standard that everyone follows, not a vague understanding that people will "be careful."

If you want a single mental model for 3.3 and 5.x together:

The AI is staff, you are the supervising lawyer, and the signature on the filing is yours.

Rule 1.5 (Fees): The Rule Nobody Reads, and the One That Surprises Clients

This is the second sleeper, and in daily practice it may be the one most lawyers are quietly getting wrong right now.

512 makes two points about fees that cut against intuition.

You generally cannot bill the client for your learning curve. Time you spend figuring out how to use a new AI tool is, in most cases, overhead, not billable client work. It is the equivalent of billing a client for the hours you spent learning Westlaw. There can be exceptions when a client specifically requests a tool that requires unusual setup, but the default is that your education is your cost.

You have to be honest about the efficiency you gain. This is the part with teeth. If a task that used to take you three hours now takes twenty minutes because the AI drafted the first pass, you cannot bill three hours. A reasonable fee under Rule 1.5 reflects the time actually and reasonably spent.

The technology that makes you faster is supposed to make the client's bill smaller, or at least it cannot inflate it. And if you intend to pass through the cost of an AI tool as an expense, you need the client's informed consent for that, framed clearly, not buried.

The reason this rule gets ignored is that it works against the lawyer's own incentives. Nobody wants to read the opinion that says "the efficiency gain belongs to the client."

But it is exactly the kind of quiet obligation that turns into a fee dispute or a bar complaint, because clients are not stupid and they can tell when a four-hour task got billed at the old rate. If you are rethinking how AI changes what you charge and what you pay for tools, our breakdown of legal research costs and what firms actually pay is the practical companion to this rule.

What firms are actually changing on Monday morning

Look at the bar opinions and firm memos that have followed 512, and a small set of operational changes show up over and over. Engagement letters now carry a separate AI-use clause that names the category of tool, the retention posture, and the specific consent ask, replacing the boilerplate "may use technology" line.

Vendor diligence packets have grown a section on training, retention, and sub-processors that procurement, not just IT, has to sign. Billing templates have been edited to strip the "research" line item from tasks where the AI did the first pass, with a separate "AI-assisted review" code that bills at a lower rate or, in some firms, not at all.

None of this is glamorous, and none of it is in 512's text. It is what 512 has actually pushed out into the day-to-day, six rules turning into six small workflow rewrites in firms that take the opinion seriously.

State Ethics Opinions That Followed 512

512 is national guidance, not a binding rule in any single state. Your state bar's opinion is what actually governs you, and by 2026 more than 25 state bars had issued formal opinions or detailed guidance on lawyer use of generative AI. The good news for compliance: they converge. Every one lands on the same core idea as 512, that the lawyer owns everything the tool produces.

A few worth knowing, because they shape what compliant practice looks like in their states:

If your state appears here, read its opinion in full before you rely on this summary, because the language that governs you is theirs, not the ABA's. Whether a formal opinion binds you at all is its own question, which we cover in are ABA formal opinions binding.

The Pattern: 512 Pushes the Burden Onto Your Workflow

Step back and look at all six rules together. The ABA's most important move in 512 is something it does not say.

It does not give you a verification threshold. It does not list approved tools. It does not define how much disclosure is enough or which tasks need consent. Over and over, it says the answer is "fact-specific" and depends on the tool and the task.

That is not the ABA being unhelpful. It is the ABA being honest that the technology changes too fast for bright lines, and that the responsibility cannot be outsourced to a rule.

So the compliance burden lands somewhere specific: on the tool you pick and the process you build around it. Two lawyers using the same opinion can have wildly different exposure depending on whether their tool retains and trains on inputs, whether it grounds answers in real authority you can open, and whether the firm has an actual policy or just hope. 512 is, in effect, a forcing function for better tool selection.

That is also why the adoption numbers matter. The 2024 ABA Legal Technology Survey (released March 2025) found that roughly 30% of lawyers report using AI, up from 11% the year before, with big firms (around 46% at firms of 100-plus lawyers) far ahead of solos (around 18%). About three in four lawyers name hallucination and accuracy as their top reason for hesitating. Only about 13% consider AI "mainstream," though some 45% expect it will be within three years (LawSites summary of the ABA survey, March 2025).

Read against 512, that hesitation is rational and slightly misdirected: the accuracy fear is real, but it is the most solvable of the six duties through tool choice. The fee and confidentiality duties are the ones quietly accruing risk while everyone watches the citation problem.

A Working Checklist (That 512 Will Not Give You)

The opinion won't hand you a checklist, so here is the operational version, built from the six duties:

• Pick a tool you can verify cheaply. Prefer AI that retrieves from real authority and links its sources, so confirming a citation takes seconds, not a re-do of the research. Statutory citations count too; sanity-check U.S. Code, CFR, and state codes against primary text rather than a memorized cite.
• Know your data posture before you type. Confirm whether the tool trains on or retains your inputs. If it is self-learning, get real informed consent first. Boilerplate does not count.
• Write the policy. Under 5.1/5.3, the firm needs a documented AI policy, a verification standard, and training. "Be careful" is not a policy.
• Verify every authority before filing. Open it. Read the passage. Confirm it holds what you claim. This is non-negotiable under 3.3.
• Fix your billing. Do not bill the learning curve. Do not bill the old hours for the new speed. Get consent before passing through tool costs.
• Communicate when it is material. Pair the 1.4 conversation with the 1.6 consent. Do not announce incidental use, but be honest with yourself about which bucket you are in.

None of this requires a law degree in AI. It requires treating the tool like staff, and treating the six rules like the connected system 512 says they are.

FAQ

What is ABA Formal Opinion 512?

ABA Formal Opinion 512 is the American Bar Association's first formal ethics opinion on generative AI, issued July 29, 2024 by the Standing Committee on Ethics and Professional Responsibility. It does not create new rules. It explains how existing Model Rules of Professional Conduct apply when a lawyer uses a generative AI tool, covering competence, confidentiality, communication, candor, supervision, and fees.

When was ABA Formal Opinion 512 issued?

The ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, titled "Generative Artificial Intelligence Tools," on July 29, 2024. It was the ABA's first formal ethics opinion addressing a lawyer's use of generative AI.

Which Model Rules does Opinion 512 address?

Six areas: competence (Rule 1.1), confidentiality (Rule 1.6, with 1.9(c) and 1.18(b)), communication (Rule 1.4), candor and meritorious claims (Rules 3.1, 3.3, and 8.4(c)), supervisory responsibilities (Rules 5.1 and 5.3), and reasonable fees (Rule 1.5). The opinion treats these as one connected system, not a checklist.

Do lawyers have to tell clients they used AI?

Not always. 512 does not require disclosing every use of AI. Disclosure becomes necessary when the client asks, when AI use is material to the representation, when it affects the basis or reasonableness of the fee, or when you need the client's informed consent to input their information under Rule 1.6. Routine, contained uses like grammar cleanup generally do not require disclosure.

Does Opinion 512 require informed consent before using AI?

Yes, in a specific situation. Before you input information relating to a representation into a self-learning AI tool, one that trains on your inputs or shares them, you generally need the client's informed consent. The opinion is explicit that boilerplate engagement-letter language about "technology including AI" does not count. The client has to understand the actual risk to their actual information.

Can lawyers bill clients for time spent using AI?

You can bill for time you actually and reasonably spend on the client's work. You generally cannot bill the client for the time you spend learning a new AI tool, the same way you would not bill for learning Westlaw. And if AI cuts a three-hour task to twenty minutes, Rule 1.5 says you bill the twenty minutes, not the old three hours. Passing through tool costs as an expense needs the client's informed consent.

Is ABA Formal Opinion 512 binding on lawyers?

No. ABA formal opinions are persuasive guidance, not law. What binds you is your own state's rules of professional conduct and any ethics opinion your state bar has issued. Many states have published their own AI guidance that tracks 512 closely. We cover the binding question in detail in are ABA formal opinions binding.

Does Opinion 512 require lawyers to verify AI output?

Yes. Under competence (Rule 1.1) and candor (Rules 3.1, 3.3, 8.4(c)), you must independently verify AI output before relying on it, and uncritical reliance on AI content is described as almost certainly malpractice. The opinion deliberately refuses to set a fixed verification standard; how much review is enough is fact-specific and depends on the tool and the task.

The "pick a tool you can verify cheaply" rule is exactly why Vaquill AI grounds its legal research in real authority with citation verification, so confirming a cite takes seconds instead of a full re-research.

Want research you can defend under 512? Start a free Vaquill AI trial, or see how citation-verified research works.

Arshita Anand

Co-Founder & CEO · Attorney

Arshita leads product and strategy at Vaquill, building the legal AI suite that solo, small-firm, and in-house US lawyers use to run a matter end to end.