What is ABA Formal Opinion 512?

A July 2024 formal opinion from the American Bar Association's Standing Committee on Ethics and Professional Responsibility. It addresses lawyers' obligations under the Model Rules of Professional Conduct when using generative AI tools, covering competence, confidentiality, communication, supervision, and reasonable fees.

Does Opinion 512 ban lawyers from using generative AI?

No. The opinion explicitly contemplates lawyers using generative AI and frames the obligations as duties around its use, not prohibitions. Confidentiality, competence, and informed-client-consent are the central considerations.

Does Opinion 512 require client consent before using AI?

It depends on whether the AI tool is self-learning (i.e. uses client information to train future outputs that other users could see) and on whether confidential information is being input. For self-learning tools or where confidential information is being shared with the tool, the opinion advises informed client consent under Rule 1.6(a).

How does on-device AI change the analysis?

On-device AI (a model running locally on the lawyer's machine, with no data transmission to a vendor) materially reduces the confidentiality concerns the opinion raises. There is no third party to disclose to, no cloud training corpus, no contractual carve-outs to negotiate. Many of the disclosure and consent obligations the opinion attaches to cloud tools don't attach the same way to architectural-privacy tools.

Is Muet compliant with ABA Opinion 512?

Muet runs entirely on the lawyer's Mac. Documents are never transmitted to any vendor. There is no cloud component to the AI inference. This architecture removes the disclosure and confidentiality concerns Opinion 512 attaches to cloud-based generative AI. Lawyers should still consult their own ethics counsel on specific circumstances; Muet is not legal advice.

Does Opinion 512 apply outside the United States?

Opinion 512 is an interpretation of the ABA Model Rules of Professional Conduct, which most US states adopt with local variations. It has no formal force outside the US, but many comparable bars (UK SRA, EU national bars, ICC Code of Conduct) apply similar duties around confidentiality, competence, and informed consent — the analytical pattern transfers.

What is the practical cost of compliance for a small firm?

For cloud-based AI, compliance involves vendor due-diligence, data-processing agreements, client disclosure language, ongoing monitoring of vendor terms changes, and supervision policies. For on-device AI, most of these vanish — there is no vendor to vet, no DPA to negotiate, no terms changes to monitor.

Where can I read Opinion 512 in full?

The opinion is published on the ABA's website (americanbar.org → Center for Professional Responsibility → Formal Opinions). The full text is approximately 8,000 words.

ABA Formal Opinion 512

What 512 actually requires.

The ABA didn’t ban AI. It also didn’t bless it.

What it did was tell every US lawyer how to use it without losing their licence.

What it is

ABA Formal Opinion 512, issued 29 July 2024, is the American Bar Association’s first formal ethics guidance on lawyers’ use of generative AI tools. It’s not binding law, since state bars set the actual rules, but it’s the framework most state bars are adopting and most malpractice carriers are now reading. If you’re a US lawyer using AI, this is the document you’re measured against.

The full text is at the ABA’s site. This page is the plain-language version of the parts that affect what tool you can use.

The six duties it touches

The opinion runs through six Model Rules and explains how each one interacts with generative AI:

Competence (Rule 1.1). You have to understand the AI tool well enough to know its limits. You can’t outsource the judgment.
Confidentiality (Rule 1.6). You owe a duty to protect client information regardless of source, and inputting client information into a tool that processes it on someone else’s servers is a disclosure.
Communication (Rule 1.4). If your use of AI is material to the representation, you may need to tell the client.
Candor to the tribunal (Rule 3.3). AI hallucinations have already caused sanctions. You sign the brief; you’re responsible for what’s in it.
Supervisory duties (Rules 5.1, 5.3). If your associates or paralegals use AI, you’re supervising them on the AI use too.
Reasonable fees (Rule 1.5). You can’t bill the client an associate hour for ten seconds of AI work.

The confidentiality part: what matters for tool choice

Rule 1.6 is where tool choice gets decided. The opinion makes three things clear:

You owe the duty regardless of the information’s source. It doesn’t matter if it’s already in a public document.
Putting client information into a generative AI tool is a disclosure to a third party, and so requires the client’sinformed consent.
Boilerplate consent buried in an engagement letter is not enough. The consent has to be specific.

The opinion is also explicit that lawyers must “know how the AI tool uses data” and put adequate safeguards in place against unauthorised disclosure to other users of the tool, vendor staff, or third parties.

What this means in practice

Most US lawyers using ChatGPT or Claude for client work are doing one of three things:

Asking specific informed consent for each matter (rare and exhausting).
Anonymising every input before pasting (the “fatigue trap” that works for the first ten redactions and collapses by the twentieth).
Relying on the vendor’s contract not to train, retain, or share while quietly hoping the rule doesn’t apply because the vendor said the right things.

The May 2025 court order in NYT v. OpenAI made the third option meaningfully worse. The court compelled OpenAI to retain deleted ChatGPT messages; Plus, Free, Team, and most API users were in scope. Whatever the vendor’s contract says, the court can override.

Where on-device AI changes the analysis

512’s confidentiality rule turns on whether a third party received the information. That’s the trigger.

When the AI runs entirely on the lawyer’s own machine, no third party receives the information. There is no processor relationship, no DPA, no sub-processor list, no vendor staff with abuse-review access, no court that can compel a vendor to produce what the vendor never had. The opinion’s confidentiality analysis still applies (you’re still responsible for competence and supervision) but the disclosure question doesn’t arise.

That doesn’t mean on-device AI satisfies every duty in 512. You still have to understand what the model produces, you still can’t file a hallucinated brief, and you still supervise associates who use it. It does mean the part of the rule that requires informed client consent for disclosure to third parties stops being a part of the conversation.

What the same logic looks like outside the US

France. The CNB’s September 2024 guidance on secret professionnel reaches the same conclusion: client information shouldn’t be communicated to AI applications, and the duty under Art. 66-5 of the 31 December 1971 law is non-waivable.
Germany. §203 StGB creates criminal liability for unauthorised disclosure of client secrets. The 2017 reform allows the use of IT providers but extends criminal liability to them, which makes a no-third-party tool meaningfully simpler than a contract-managed cloud one.
United Kingdom. The SRA Code Section 6 maintains confidentiality as a fundamental duty. The same trigger logic applies: a tool that doesn’t involve disclosure doesn’t engage the same analysis.
EU more broadly. GDPR Art. 28 (data processors), Arts. 44–49 (transfers under Schrems II), and the EU AI Act all turn on whether a processor or controller relationship exists with a third party. With on-device AI, none does.

What Muet is, in this context

Muet is a Mac app that reads, answers, and edits your documents entirely on your Mac. No cloud processing, no prompts logged, no third-party vendor in the loop for the work itself. It exists for the part of legal practice that 512’s confidentiality analysis otherwise complicates.

It is not legal advice and we are not your lawyer. The opinion is the opinion; this page is one reading of it; your jurisdiction’s bar may have its own gloss; talk to your malpractice carrier before relying on any single article.

Pricing

$390per Mac, per year

30-day money-back guarantee. Apple Silicon (M1 or newer), macOS 14 or newer. One licence per device, easy to expense.

Get Muet →