Somewhere in your engagement letter there is probably a sentence saying the firm “may use artificial intelligence tools in the course of the representation”. ABA Formal Opinion 512 says that sentence does nothing: “Merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient.”
That is the most useful thing about the opinion: it is specific where most AI guidance is vague, on what “competence” means for a tool you did not build, what informed consent must contain, and how many minutes you may bill when a draft that used to take three hours now takes one. Here it is, duty by duty, with the opinion’s own examples and the state opinions built on it since.
What Opinion 512 is and is not
The ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, “Generative Artificial Intelligence Tools”, on 29 July 2024. It applies ten existing Model Rules: 1.1 (competence), 1.4 (communication), 1.5 (fees), 1.6 (confidentiality), 1.9© and 1.18(b) (former and prospective clients), 3.1 and 3.3 (candour) and 5.1 and 5.3 (supervision).
It is not a rule: ABA opinions interpret the Model Rules, which bind nobody until a state adopts them. Nor is it a prohibition; it says “it is conceivable that lawyers will eventually have to use them to competently complete certain tasks for clients.” Read it as the floor; the state bar AI ethics opinions map shows where your state sits higher.
Competence: “reasonable understanding”, not expertise
Lawyers “need not become GAI experts” but “must have a reasonable understanding of the capabilities and limitations” of the specific tool they use. That builds on Comment 8 to Model Rule 1.1, the 2012 technology-competence amendment adopted by 40 states plus DC and Puerto Rico. The opinion warns against “uncritical reliance on content created by a GAI tool” and makes the required review a sliding scale that “will necessarily depend on the GAI tool and the specific task that it performs”.
The profession knows it has not met this duty: in the State Bar of Texas’s 2026 survey of 1,553 lawyers, only 9% felt “extremely confident” evaluating AI output, and 67% supported required AI-ethics CLE.
Confidentiality: self-learning tools and informed consent
Rule 1.6 is where the opinion does its real work. “Before lawyers input information relating to the representation of a client into a GAI tool, they must evaluate the risks that the information will be disclosed to or accessed by others outside the firm.” Lawyers should “read and understand the Terms of Use, privacy policy, and related contractual terms” or get help doing so.
Then the trigger. Self-learning tools “by their very nature raise the risk that information relating to one client’s representation may be disclosed improperly, even if the tool is used exclusively by lawyers at the same firm”, so “a client’s informed consent is required prior to inputting information relating to the representation into such a GAI tool”.
What “self-learning” means on the tools you actually use
The opinion speaks of tools in the abstract, so here is the translation into product tiers as of September 2026. ChatGPT Free, Plus and Pro train on conversations unless you switch off “Improve the model for everyone”; Claude Free, Pro and Max have trained by default since 28 August 2025, with five-year retention for opted-in users; consumer Gemini, consumer Copilot and Perplexity Free and Pro do the same. The business, enterprise and API tiers of each, Copilot with enterprise data protection and the legal platforms do not train by default. The consent trigger usually turns on the tier, not the vendor; is ChatGPT confidential for lawyers? has the tier-by-tier table.
Why boilerplate engagement-letter consent fails
Informed consent has content. The client needs “the lawyer’s best judgment about why the GAI tool is being used”, “the extent of and specific information about the risk”, “the ways in which others might use the information against the client’s interests” and “a clear explanation of the GAI tool’s benefits”. A clause saying “the firm may use AI tools” contains none of the four, which is why it is “not sufficient”.
Draft a consent paragraph for our engagement letter, for a [law firm type] in [state], with the four elements ABA Formal Opinion 512 requires: (1) why we use the tools (enterprise tools with contractual no-training and retention terms; no consumer tools for client information); (2) the specific risks (error, retention, disclosure to the provider, loss of privilege where terms allow third-party access); (3) how others could use the information against the client; (4) the benefits. Add the client's right to instruct otherwise and one sentence on billing (actual time only; no charge for learning tools). Plain English, under 200 words.Communication: when to tell the client
Rule 1.4 disclosure is mandatory in three situations: when the client asks, when client information will be input (because consent is needed), and when the use is “relevant to the basis or reasonableness of a lawyer’s fee”. Outside those the question is “factually specific”, and the states have answered it differently: North Carolina says “generally, a lawyer need not inform her client that she is using an AI tool to complete ordinary tasks”, while Pennsylvania leans to being “transparent with clients, colleagues, and the courts”. Do lawyers have to disclose AI use? takes it state by state.
Supervision: Rules 5.1 and 5.3 for AI
“Managerial lawyers must establish clear policies regarding the law firm’s permissible use of GAI, and supervisory lawyers must make reasonable efforts to ensure that the firm’s lawyers and nonlawyers comply with their professional obligations when using GAI tools.” North Carolina adds that Rule 5.3 reaches “third-party software companies”.
The 8am 2026 Legal Industry Report found 69% of legal professionals personally using generative AI for work, while only 9% work at a firm with a written, enforced policy. And a policy alone is not supervision. Butler Snow had one from June 2023; in Johnson v. Dunn (23 July 2025) Judge Manasco still disqualified three partners and referred them to the Alabama bar, noting they “benefitted from repeated warnings, internal controls, and firm policies about the dangers of AI misuse… And yet here we are.” The law firm AI policy template shows what a policy that functions as a control looks like.
Candour and meritorious claims
Every citation, quotation and assertion in a filing is the signing lawyer’s, whoever or whatever drafted it. Damien Charlotin’s database listed 2,039 decisions involving hallucinated material on 12 September 2026, 811 involving lawyers, and the Fifth Circuit’s line is the one courts repeat: “‘I used AI’ will not be an excuse for an otherwise sanctionable offense.” Verification means opening the authority in a database and reading the pinpoint; asking the model whether its citations are real is what Steven Schwartz did in Mata v. Avianca.
Here is a workflow we run: <workflow>[e.g. paralegal uploads the counterparty's NDA to Claude Team, runs our playbook review, associate edits the redline, partner sends it]</workflow>. For each ABA Formal Opinion 512 duty (competence, confidentiality, communication and consent, supervision, candour, fees) state: the step where the duty bites; what the opinion requires there, quoting <opinion>...</opinion>; the evidence we would need to show compliance (tier and contract terms, consent language, review log, time entry); and the most likely failure. Output as a table, then three lines for the practice-group head. Do not add requirements the opinion does not contain.Fees: the 15-minute example and no charging to learn
When billing hourly, lawyers may “only bill for their actual time”, and: “A fee charged for which little or no work was performed is an unreasonable fee.” The worked example: a lawyer who “expends 15 minutes to input the relevant information into the program… may charge for that time as well as for the time necessary to review the resulting draft”, but “in most circumstances, the lawyer cannot charge a client for learning how to work a GAI tool”. Matter-specific costs pass through at actual cost with no surcharge; the state opinions treat a general subscription as overhead.
North Carolina’s 2024 Formal Ethics Opinion 1 turns this into arithmetic. An estate planner billing $300 an hour used to spend three hours on standard documents and now spends one. “No, Lawyer may not bill a client for three hours of work when only one hour of work was actually experienced.” She may instead complete “more work for more clients”, or move to a flat fee, permitted “provided the flat fee charged is not clearly excessive and the client consents”.
| Time or cost | Billable? |
|---|---|
| 15 minutes writing the prompt and inputting facts | Yes |
| 60 minutes reviewing and correcting the draft | Yes |
| The two hours the tool saved | No |
| Learning how the tool works | No, “in most circumstances” |
| The firm’s subscription | No; overhead (California: “similar to library maintenance”) |
| A matter-specific AI charge | At actual cost, no mark-up, with consent, preferably written |
| A flat fee for a matter type | Yes, if not clearly excessive and the client consents |
Review these draft time entries <entries>...</entries> for a matter in which we used [tool, tier] for [tasks]. For each entry confirm it records actual time spent prompting, inputting facts or reviewing output; flag any entry that bills time the tool saved, reconstructed "equivalent" time or time spent learning the tool; flag block billing; and flag any AI disbursement that is not a matter-specific charge at actual cost with recorded client consent. Do not change the hours; list the entries needing the timekeeper's attention with a one-line reason each.The compliance checklist
One page, six duties, one piece of evidence per duty. If the evidence column is empty, the duty is not met, whatever the policy says.
| Duty | Opinion 512 requirement | Evidence the firm keeps |
|---|---|---|
| Competence (1.1) | “Reasonable understanding” of each tool’s capabilities and limits | Approved-tool list with tier; training record; written verification step per task type |
| Confidentiality (1.6, 1.9©, 1.18(b)) | Risk evaluation before input; no client data in self-learning tools without informed consent | Vendor file: no-training clause, retention, DPA; anonymisation procedure |
| Communication (1.4) | Disclose when asked, when consent is required, or when fees are affected | Consent paragraph with the four elements; note of client instructions |
| Supervision (5.1, 5.3) | Clear policies; reasonable efforts; training | Policy naming tools and data tiers; acknowledgement on hire; quarterly review |
| Candour (3.1, 3.3) | Verify every citation, quotation and assertion | Pre-filing verification log: who checked what, where, when |
| Fees (1.5) | Actual time only; no charge for learning; costs at cost | Time-entry guidance; disbursement policy; flat-fee scoping notes |
How states have built on 512
Several states have raised the floor. Florida’s Opinion 24-1 (19 January 2024) says a lawyer may not delegate “any act that could constitute the practice of law such as the negotiation of claims”. New York City’s 2024-5 (7 August 2024) says that even with consent lawyers should “avoid entering details that can be used to identify the client”. California’s rewritten COPRAC guidance (14 May 2026) goes furthest: “Lawyers must not permit AI systems to autonomously file documents, communicate with the court, or make representations on the lawyer’s behalf”, and “reasonable efforts require more than reliance on generalized marketing assurances”.
Where to go next: the cross-border view of the same duties is in AI ethics rules for lawyers by jurisdiction; the CLE requirements behind Rule 1.1 are in AI CLE and technology CLE requirements; the ethics and regulation cluster and the prompt library hold the rest. A “reasonable understanding” of the tools is hard to get from a slide deck; AI Lab for Lawyers builds it in four live sessions on your own anonymised documents, with each workflow mapped to the duties above.
Frequently asked questions
What does ABA Opinion 512 require?
It applies Model Rules 1.1, 1.4, 1.5, 1.6, 1.9(c), 1.18(b), 3.1, 3.3, 5.1 and 5.3 to generative AI. Lawyers must understand a tool's capabilities and limitations, evaluate confidentiality risk before inputting client information, obtain informed consent for self-learning tools, tell clients when asked or when fees are affected, verify every citation and assertion, supervise staff and vendors, and bill only actual time. It was issued on 29 July 2024.
Do I need client consent to use ChatGPT?
Under Opinion 512, yes if client information will be entered into a tool that learns from inputs, which consumer ChatGPT tiers do by default unless the training toggle is off. ChatGPT Business, Enterprise and API tiers do not train by default, so the consent trigger is usually the tier, not the product. Consent must be informed: the client is told why the tool is used, the specific risks and the benefits. A boilerplate clause is not enough.
Can I bill for time saved by AI?
No. Opinion 512 says lawyers billing hourly 'may only bill for their actual time' and that 'a fee charged for which little or no work was performed is an unreasonable fee'. You may bill the minutes spent prompting and the time spent reviewing the output, but not the hours the tool replaced and not the time spent learning the tool. North Carolina's 2024 FEO 1 puts it plainly: three hours of drafting that now takes one is billed as one.
Is Opinion 512 binding?
Not by itself. ABA formal opinions interpret the Model Rules, which bind nobody until a state adopts them, and they are advisory even then. In practice Opinion 512 is the reference text: the later state opinions from North Carolina, Texas and California cover the same duties, and its competence and consent standards are the ones ethics counsel start from. Treat it as the standard you will be measured against unless your state has written something stricter.
What is a self-learning AI tool?
A tool that uses what you type to train or improve the model, so information from one matter can surface elsewhere. Opinion 512 says such tools 'by their very nature raise the risk that information relating to one client's representation may be disclosed improperly, even if the tool is used exclusively by lawyers at the same firm'. Consumer ChatGPT, Claude, Gemini, Copilot and Perplexity tiers train by default; business, enterprise and API tiers and the legal platforms do not.