If your firm uses generative AI for anything client-related, from summarizing a production to drafting a section of a brief, ABA Formal Opinion 512 is the document your AI policy should be built around. Issued July 29, 2024 by the ABA Standing Committee on Ethics and Professional Responsibility, it does not create new rules. It applies the Model Rules your lawyers already follow to a new kind of tool, and it is specific enough to turn into written procedures.
Two caveats up front. ABA formal opinions are advisory; your state’s version of the rules, and any guidance from your state bar or local courts, control. Many bars have also published their own AI guidance. Read the opinion itself, then read what applies where you practice. What follows translates the opinion’s main areas into steps a litigation team can actually take.
What is ABA Formal Opinion 512?
Opinion 512 is the ABA’s formal ethics opinion on generative AI: tools that create new content such as text, summaries and analysis in response to prompts. It walks through how the duties of competence, confidentiality, communication, candor, supervision and reasonable fees apply when lawyers use those tools.
The committee was explicit that the technology is changing quickly and that guidance may need to be updated. The opinion does not approve or ban any product. It describes a process: understand the tool, protect client information, tell clients what they need to know, check the output, supervise the people using it and bill fairly. A policy built on that process will outlast the next product cycle. A policy built on a list of approved brand names will not.
What does Opinion 512 require for competence?
Under Rule 1.1, a lawyer using generative AI needs a reasonable understanding of the capabilities and limitations of the specific tool in use. Nobody has to become an AI engineer, and the opinion recognizes that this understanding can come from study or from consulting people who have the expertise.
The core of the competence discussion is verification. Generative tools can produce fluent, confident output that is wrong, and the opinion warns that uncritical reliance on that output can fall short of competent representation. How much independent checking is appropriate depends on the tool and the task. Professional judgment stays with the lawyer. The opinion also leaves open the possibility that, as these tools mature, competent practice could eventually involve using them for some tasks, so “we don’t use AI” is a position to revisit periodically rather than a permanent safe harbor.
Scaling verification to the task
| Task | Risk if the output is wrong | Reasonable verification |
|---|---|---|
| Brainstorming deposition topics or issue lists | Low: a lawyer filters every idea | Lawyer judgment on what to keep |
| Summarizing a large document production | Medium: a missed or distorted fact shapes strategy | Spot-check against source pages; read anything you rely on |
| Extracting text from screenshots or scans for exhibits | High: wrong words or dates reach the court | Compare every extracted line to the original image |
| Drafting brief sections with citations | High: fabricated or misstated authority | Pull and read every cited authority before filing |
When do you need client consent to use AI?
Rule 1.6 requires a lawyer to assess, before entering information relating to the representation into an AI tool, the risk that the information will be disclosed to or accessed by others inside or outside the firm. For self-learning tools, meaning tools that may use inputs to train or improve the system, the opinion concludes that the client’s informed consent is required first.
Three details from this section belong in your written policy:
- Boilerplate is not enough. A generic clause in an engagement letter does not supply informed consent. The client needs an explanation of the particular tool’s benefits and risks.
- The risk exists inside the firm, too. A tool that learns from one matter could surface that information to colleagues working on another, including people screened from the matter.
- Someone has to read the terms. Lawyers should review the tool’s terms of use, privacy policy and related contract terms, or rely on a qualified person who has, to learn who can access inputs and whether they are retained or used for training.
Sample wording for a consent conversation
“We’d like to use [tool] to help [summarize your text messages / prepare a first draft of the timeline]. It speeds up the work, and a member of our team checks everything it produces against the original documents. The provider’s terms say [plain-language description of retention and training]. In practical terms, that means [describe the risk]. You can say no, and we will do this work without it. Do you agree to our using it for this purpose?”
The same diligence applies to any cloud system that holds client files. Our guide to client confidentiality and cloud software covers the vendor questions in detail.
Do you have to tell clients you are using AI?
Not for every use. Under Rule 1.4, the opinion treats disclosure as dependent on the facts rather than automatic.
Disclosure is called for when the client asks how the work is being done or whether AI is involved, when informed consent is required under Rule 1.6, when the tool’s output will influence a significant decision in the representation (an assessment of likely outcomes, for example), and when AI use bears on the basis or reasonableness of the fee. If an engagement agreement or the client’s outside counsel guidelines address AI, those terms govern too. Routine uses where a lawyer independently does the substantive work, such as brainstorming, are less likely to require disclosure on their own.
What does the opinion say about candor and checking output?
Rules 3.1, 3.3 and 8.4(c) make the lawyer responsible for what is filed with or said to a tribunal, regardless of who or what drafted it. The opinion points to the risk of fabricated citations and misstated law and expects lawyers to review AI output for accuracy before relying on it.
It also notes that courts may impose their own requirements, such as disclosing AI use in filings, so checking the assigned judge’s standing orders and the local rules belongs on every filing checklist. Make verification a defined step with a named owner. Our guide to preventing AI hallucinations in legal work lays out a full verification protocol.
What are the supervision duties under Rules 5.1 and 5.3?
Lawyers with managerial authority must put clear policies in place on permissible AI use, and supervising lawyers must make reasonable efforts to ensure that associates, paralegals and other staff follow them. The opinion also expects training on how the tools work, their ethical risks and the firm’s rules.
When AI comes from an outside vendor, the opinion draws on the kind of due diligence the ABA has long expected for outsourcing and cloud services: understand the vendor’s security and confidentiality practices, review the contract, and confirm what happens to your data. For staff-level detail, see Rule 5.3 and supervising AI tools and our guide to using AI ethically as a paralegal.
How should firms bill for AI-assisted work?
Rule 1.5 requires reasonable fees and expenses, and the opinion applies that standard to AI in concrete terms. Under hourly billing, a lawyer bills for time actually spent, which can include writing prompts and reviewing the output, but not the hours the tool saved.
- Flat and contingent fees still have to be reasonable under the Rule 1.5(a) factors. A fee set on the assumption of pre-AI effort may deserve a second look when a tool makes the work much faster.
- Tool costs. Tools that function as general overhead, much like word processing, generally are not billed to clients as expenses without an agreement. A tool used for a specific client’s matter may be passed through at actual cost if the client is told in advance.
- Learning time. Time spent learning a tool the lawyer will use across matters is part of maintaining competence and generally is not billable. The narrow exception is a client who asks for a specific tool the lawyer does not already use.
An Opinion 512 implementation checklist
Use this as the agenda for your firm’s AI policy meeting:
- Inventory every AI tool in use, including AI features built into software you already pay for.
- For each tool, document whether inputs are retained, who can access them and whether they train the model.
- Classify each tool: approved for client information, approved for non-confidential use only, or prohibited.
- Write a consent script and a way to record informed consent where Rule 1.6 requires it.
- Define verification steps by task type, with a named owner for every filing.
- Keep AI output separate from original evidence, and require a person to confirm it before it enters the record.
- Add court and judge-specific AI disclosure requirements to the filing checklist.
- Update billing guidance for hourly work, flat fees and expenses.
- Train lawyers and staff, and repeat the training when tools or rules change.
- Assign someone to track your state bar’s guidance and revisit the policy at least once a year.
Software can make some of these controls structural rather than aspirational. Evidence Helper, for instance, keeps AI-extracted text from screenshots separate from the hashed original and stores a reviewer’s corrections separately, so the human check is part of the workflow instead of something people have to remember (see how the extraction review works). Whatever tools you choose, Opinion 512 comes down to three questions: does a person verify the output, is client information protected, and can you explain both to the client? When the answer turns on your particular practice or jurisdiction, your state bar’s ethics resources or a licensed ethics attorney where you practice can help.
