ABA Formal Opinion 512 was issued on 29 July 2024 and remains the reference point for how United States lawyers may use generative artificial intelligence. It is not long, and it is not especially technical. It is also widely cited by people who have clearly not read it, usually to support a position it does not take.
What follows is what the opinion actually says, and what it means in practice.
The framing: no new rules
The opinion does not create a special regime for AI. It applies existing Model Rules to a new tool. That framing matters, because it means the analysis you already apply to outsourcing, to cloud storage, and to non-lawyer assistance is the analysis that applies here.
The rules it works through are competence (Model Rule 1.1), confidentiality (1.6), communication (1.4), and fees (1.5).
Competence: you have to understand the tool
Rule 1.1 requires competent representation, and its commentary already required lawyers to keep abreast of the benefits and risks of relevant technology. Opinion 512 applies that directly: a lawyer using a generative tool must understand what it does and where it fails.
In practice this sets a floor rather than a ceiling. You are not required to understand model architecture. You are required to know that these systems produce fluent output that can be wrong, that they can fabricate citations, and that their output is a draft rather than a result. A lawyer who does not know that a general-purpose chatbot invents case law is not meeting the standard, and the volume of sanctions cases since 2024 has made that difficult to argue.
The duty is also ongoing. These tools change materially every few months. Competence acquired in 2024 does not automatically carry.
Confidentiality: the question is what happens to the input
This is the operative section for most firms. Rule 1.6 protects information relating to the representation, and the opinion treats entering that information into a generative tool as engaging the duty. Protection extends to prospective and former clients as well.
The analysis turns on what the tool does with what you give it. A self-contained system that does not retain or train on inputs sits differently from a consumer product whose terms permit the provider to use submissions to improve its models.
Where disclosure to the provider would occur, the opinion looks to informed client consent. Informed means the client understands what is being disclosed and to whom. A line buried in an engagement letter saying the firm “may use technology” is not informed consent to uploading their confidential documents to a third party.
The practical sequence is straightforward. Read the terms of service and the data processing terms, not the marketing page. Establish whether inputs train the model. Establish retention and deletion. Establish who at the vendor can access the data. Then decide whether consent is needed, and if it is, ask for it specifically.
Communication: telling the client
Rule 1.4 requires reasonable consultation about the means used to accomplish the client’s objectives. The opinion does not impose a blanket duty to announce every use of AI, which is the misreading that circulates most often.
Disclosure is required where use of the tool is material to the representation, where confidential information would be disclosed to the provider, or where the client has asked. Many sophisticated clients have now asked, often through outside counsel guidelines with specific AI provisions. Those guidelines are contractual and they govern regardless of what the ethics opinion permits.
Fees: the section that catches people out
Rule 1.5 prohibits unreasonable fees. Opinion 512 draws two clear lines.
You may bill for time actually spent working with the tool. Drafting the prompt, assembling the inputs, and reviewing and correcting the output are all legal work and are billable.
You may not, in most circumstances, bill a client for learning how to use the tool. That is a firm overhead, in the same way that learning your document management system is.
The harder question sits underneath. If a task that took four hours now takes one, a lawyer billing hourly bills one hour. The opinion does not require you to pass on efficiency gains as a discount, but it does mean you cannot bill the four.
This is where the profession’s practice and its economics are visibly out of step. Survey data from 2026 shows 86 per cent of solo firms and 78 per cent of small firms have not adjusted pricing despite adopting these tools. Some of that is inertia. Some of it is firms quietly capturing the gain. Clients with procurement functions have noticed.
What the opinion does not do
It does not approve or prohibit any particular product. It does not set a disclosure requirement for court filings, which is a matter for individual judges and their standing orders. It does not address the European position, where the EU AI Act creates a separate set of obligations. And it is not binding anywhere on its own, since it is guidance on Model Rules that each state adopts and interprets for itself. Several state bars have issued their own opinions, and some go further.
What to do with it
Four things, in order. Read the data terms for every generative tool your firm uses, and write down the answer on training, retention and access. Decide, per tool and per matter type, whether client consent is required, and build the ask into your engagement process rather than handling it case by case. Set a billing position on tool learning time and communicate it internally. And write it all down, because more than half of firms still have no AI policy, and the opinion is considerably easier to comply with than to explain non-compliance with afterwards.
See also our practical structure for writing a firm AI policy and the record of what happens when verification fails.
Where the legal industry reads first.
Enjoyed this article? Get the biggest legal industry updates, deals, appointments, insights and expert interviews in your inbox, free.
No spam. Unsubscribe anytime.
