Do You Have to Tell Your Client You Used AI? An Informed-Consent Walkthrough Under ABA Opinion 512

The short answer is that it depends on the tool and the data

Opinion 512 says a lawyer may need informed consent before putting client information into a self-learning AI tool. "May" is doing real work. The answer depends on the tool and the data, not on a fixed rule.

There is no bright-line rule that a lawyer must announce every use of AI to a client. ABA Formal Opinion 512, issued in July 2024, frames AI use as implicating several professional duties at once, and on the confidentiality duty it says that before inputting information relating to the representation into a self-learning generative AI tool, a lawyer may need the client's informed consent ^[1]. The word that carries the weight there is "may." The opinion is guidance, and the analysis is fact-specific rather than automatic.

That framing matters because two lawyers can reach different, defensible answers for the same task depending on the tools they use and the data they handle. A lawyer summarizing a public appellate opinion in a consumer chatbot is in a different position from a lawyer pasting a client's confidential settlement terms into that same tool. A lawyer using an enterprise system with a contractual no-training and zero-retention commitment is in a different position again. The question is not simply "did you use AI," it is "what did the tool do with what you gave it, and was any of it confidential."

So the useful move is not to hunt for a universal rule that does not exist. It is to learn which facts push a specific use toward needing consent and which facts push it away, then run that analysis honestly on each matter. The rest of this article does that, keeps the ethical confidentiality duty separate from the evidentiary privilege question throughout, and ends where any responsible version of this has to end, which is with your own counsel and your jurisdiction's adopted rules.

"Do I have to tell my client" is really two questions

The question sounds like one thing and is actually two. The first is ethical: does using an AI tool on this matter trigger a professional duty to inform the client and obtain consent? That question lives under the confidentiality duty in Model Rule 1.6, which is where Opinion 512 grounds its AI guidance ^[1][2]. The second is evidentiary: does routing client material through an AI system risk waiving attorney-client privilege? That question is decided by courts, case by case, and it is genuinely unsettled.

These two questions run on separate tracks, and a clean answer to one does not answer the other. A lawyer can satisfy every consent and competence expectation in Opinion 512 and still face an open question about whether a specific disclosure affected privilege. Conversely, a use that raises no privilege issue at all can still require consent under the confidentiality duty if it puts client-identifiable information into a self-learning tool. Treating "I used enterprise AI" as if it disposes of both questions is a category error.

This article is about the first question, the confidentiality and informed-consent duty that Opinion 512 addresses ^[1]. It touches the privilege question only to keep it visibly separate, because conflating the two is the single most common way firms talk themselves into a false sense of safety. For a deeper treatment of the evidentiary side, see our walkthrough on AI and attorney-client privilege and the isolation question in proving privileged data never reached a model.

Confidentiality and privilege are different duties, enforced by different authorities at different times. This article is about the confidentiality and consent question, not about when privilege is waived.

What ABA Formal Opinion 512 actually says about consent

Opinion 512 does not treat AI as a single problem with a single fix. It identifies a set of Model Rules that generative AI use can implicate, including competence under Rule 1.1, communication under Rule 1.4, confidentiality under Rule 1.6, candor to the tribunal under Rule 3.3, supervision under Rules 5.1 and 5.3, and reasonable fees under Rule 1.5 ^[1]. The consent question sits inside the confidentiality duty, but it does not stand alone, because a consent a lawyer cannot competently explain is not a meaningful consent.

On confidentiality specifically, the opinion's structure is conditional rather than absolute. Before inputting information relating to the representation into a self-learning generative AI tool, a lawyer may need the client's informed consent ^[1]. That framing does two things at once. It flags a real duty that boilerplate does not satisfy, and it declines to convert that duty into a mechanical rule that fires on every AI interaction regardless of the facts. The lawyer still has to do the analysis.

Reading the opinion as a single consent checkbox misses how it works. It also ties competence and supervision to the same use, so that the responsible lawyer understands the tool well enough to judge its risks, and the firm oversees how associates and staff deploy it ^[1]. Those obligations travel together. Consent without competence produces a disclosure the lawyer cannot honestly stand behind, and competence without supervision leaves the tool being used in ways no one is watching.

1. ConfidentialityUnder Rule 1.6, a lawyer may need the client's informed consent before inputting information relating to the representation into a self-learning generative AI tool ^[1].
2. CompetenceUnder Rule 1.1, the responsible lawyer should understand the tool well enough to judge how it handles inputs and what the risks are ^[1].
3. SupervisionUnder Rules 5.1 and 5.3, the firm oversees how lawyers and nonlawyer staff use these tools, so use sits inside policy rather than being ad hoc ^[1].

Opinion 512 is ethics guidance, not a bright-line rule. It says a lawyer may need consent, and it leaves the determination fact-specific.

Why confidentiality here is not the same as evidentiary privilege

Consent under Opinion 512 addresses the ethical confidentiality duty. It is not a determination that privilege is preserved. Whether AI use waives privilege is a separate, unsettled question for courts.

It is worth stating plainly, because the distinction does real work throughout this analysis. The duty Opinion 512 addresses is the ethical duty of confidentiality under Model Rule 1.6, which covers all information relating to a representation, whether or not anything is ever litigated ^[1][2]. Evidentiary privilege is a narrower protection that can keep certain lawyer-client communications out of a proceeding, and it can be waived by disclosure to a third party.

The confidentiality duty is enforced continuously, by bar regulators and by the client relationship, and it applies to every matter a lawyer touches. Privilege only becomes a live issue when someone tries to compel a communication in a proceeding and the lawyer resists on the ground that it is protected. So the day-to-day discipline around AI is a confidentiality discipline, and privilege is a separate risk that can surface later, on a different timeline, in front of a different decision-maker.

Whether using an AI tool waives privilege is an unsettled legal question that courts decide on the facts. Opinion 512 does not resolve it, and this article does not claim to. When you see consent discussed below, read it as the confidentiality duty under Rule 1.6, not as a control over privilege waiver. Keeping these two apart is how a firm avoids answering the easier ethical question and quietly assuming the harder evidentiary one took care of itself.

The factors that move the needle toward or away from consent

Because the analysis is fact-specific, the practical work is knowing which facts push a given AI use toward needing consent and which push it away. None of these factors is dispositive on its own, and a real matter usually presents a mix. The value of laying them out is that it lets a lawyer see, before touching a tool, roughly where a proposed use sits on the spectrum and how much care it calls for.

The table below summarizes the factors drawn from Opinion 512's confidentiality analysis ^[1]. It is a working reference, not a scoring system, and it does not replace judgment on a specific matter in a specific jurisdiction. Read it as a way to structure the conversation you should be having with your own counsel, not as a formula that produces an answer by itself.

No single factor decides the question. A use can sit in the left column on one row and the right column on another. The mix is what a lawyer has to weigh.

A decision walkthrough for a single matter

Rather than reach for a universal rule, it helps to run the same short sequence on each matter before client material reaches a tool. The steps below map to the factors above and keep the confidentiality question separate from the privilege question, so that neither gets skipped. This is a way to make the fact-specific analysis routine, not a substitute for judgment or for your jurisdiction's adopted rules.

Work the sequence in order, and stop to get consent, or to change tools, at the first point where the honest answer says you should. The goal is not to reach a predetermined outcome. It is to make sure the decision is deliberate and defensible, and that you can later explain why you handled a given matter the way you did.

1. Classify the dataDecide whether the input includes information relating to the representation, and how sensitive it is, before any tool touches it.
2. Check the toolDetermine whether the tool is self-learning and whether it carries a contractual no-training and zero-retention commitment with appropriate safeguards ^[1].
3. Check the engagement letterAsk whether the client has already given informed consent covering this specific kind of AI use, or whether only boilerplate is in place.
4. Decide on consentIf confidential input is going into a self-learning tool without prior informed consent, obtain informed, specific consent before proceeding, per Opinion 512 ^[1].
5. Run the privilege question separatelyAsk, as a distinct matter, whether the disclosure could affect privilege, recognizing that this is an unsettled question for courts, not something consent resolves.
6. Record what happenedKeep a record of which tool saw which data, whether it was isolated, and that consent was captured, so you can show your handling rather than merely assert it.

This walkthrough is a structuring aid, not legal advice. Apply it with your own counsel and your jurisdiction's rules on the specific matter in front of you.

How the engagement letter changes the analysis

One factor deserves its own treatment because it can shift the answer more than any other, which is what the engagement letter already says. Opinion 512 is clear that generic, buried language is not enough to carry an informed-consent obligation, because consent has to be informed and specific to the risk ^[1]. A single line stating that the firm "may use technology, including artificial intelligence" does not, on its own, do the work of informing a client about a self-learning tool that could train on their confidential information.

That said, an engagement letter can meaningfully change the analysis when it addresses AI use with enough specificity that the client genuinely understood what they were agreeing to. If a client has been told, in terms they could act on, how the firm uses AI tools, what categories of information those tools may see, and what the safeguards are, then the consent factor moves toward the away column for uses that fall inside what was described. The letter does not become a blanket permission slip, but it can cover a defined scope.

The practical implication is that AI consent is better handled as a considered part of client intake than as a scramble in the middle of a matter. A firm that addresses AI use specifically at engagement, and revisits it when its tools or practices change, is running the confidentiality analysis by design rather than by luck. For how the confidentiality duty plays out across a representation, see our overview of client confidentiality and AI.

Boilerplate does not satisfy Opinion 512's informed-consent standard. Specific, informed language in an engagement letter can cover a defined scope of AI use, but it is not a blanket waiver.

Where state guidance varies

ABA formal opinions are influential but not binding. They are guidance that states adopt, adapt, or decline, and several state bars have issued their own guidance on lawyers using generative AI. That means the answer to "do I have to tell my client" can differ depending on the jurisdiction whose rules govern your practice, and checking your own state's guidance is not optional.

The Florida Bar has issued an ethics opinion addressing lawyers' use of generative AI, including how AI's effect on billing and costs should be disclosed to clients ^[3]. That is a different emphasis from a pure confidentiality-consent framing, and it shows how state guidance can foreground concerns the ABA opinion treats more briefly. In Texas, the State Bar has issued guidance emphasizing human oversight of AI in legal work ^[4]. The specifics vary, and this article does not try to catalog every state or assert holdings beyond what these sources say.

The safe posture is to treat the ABA opinion as the baseline framework and then read it together with your jurisdiction's own guidance, which may add requirements around disclosure, billing, or oversight. A national survey of state approaches can help you see where your jurisdiction sits ^[5], but the controlling source is always the rule your licensing authority has actually adopted, interpreted by your own counsel.

• ABA Formal Opinion 512 is influential guidance, not a binding rule that states must follow ^[1].
• Florida's ethics opinion addresses AI's impact on billing and cost disclosure to clients ^[3].
• Texas guidance emphasizes human oversight of AI in legal work ^[4].
• Check your own jurisdiction's adopted rules and opinions, because requirements differ by state.

What "informed" consent actually requires you to explain

Informed consent means the client understood the risk in plain terms: what the tool is, whether it learns from inputs, what information is involved, and how the firm limits exposure. A record of that conversation is what makes it defensible later.

If a matter lands in the territory where consent is needed, the next question is what a client actually has to be told for that consent to count. Opinion 512 ties consent to competence for a reason: a lawyer cannot obtain informed consent about a tool the lawyer does not understand ^[1]. The client needs enough of an explanation to grasp the risk they are being asked to accept, in plain terms, not a technical disclosure they cannot parse.

In practice, that means being able to tell the client, at a level they can act on, what the tool is, whether it is self-learning and could train on their information, what category of their information would be involved, and what the firm is doing to limit the exposure. If the firm is using an enterprise tool with a no-training and zero-retention commitment, saying so is part of informing the client, because it is directly relevant to the risk. The explanation should match the actual use, not a generic description of AI.

This is also where records matter. A consent that was informed and specific is only useful later if the firm can show what the client was told and that they agreed. Keeping a clear record of the consent conversation, tied to the specific use it covered, turns a professional obligation into something the firm can demonstrate rather than merely recall. That record is separate from, and does not resolve, any later privilege question.

Attesting that the consent step happened and privileged material was isolated

Whatever a firm decides on a given matter, there is a gap between doing the right thing and being able to show it later. A firm can obtain informed consent, choose an enterprise tool, and isolate the most sensitive material, and still, months later, be unable to demonstrate that any of it happened the way it remembers. Consent conversations fade into recollection, and "we isolated that data" becomes an assertion rather than a record.

This is the narrow thing verification tooling can help with, and it is worth being precise about what it does and does not claim. RankShield Legal's approach can attest that the consent step occurred and that privileged or confidential material was isolated, whether by being withheld, redacted, tokenized, or handled by a local model, binding together the interaction, the approved tool, the firm's policy, and the captured consent into a record that can be verified independently later. That attests to architecture and to consent. It does not attest to a privilege outcome.

The honest boundary is the same one this whole article has held. An attestation can show that isolation functioned as designed and that consent was captured. It cannot decide whether privilege was waived, because that is a legal determination a court makes, not a control any vendor can enforce. What it changes is not the law. It changes whether a firm can show what it did, at the moment the question is asked. The mechanics of that isolation are covered in our note on privilege isolation.

Attestation evidences that consent was captured and that material was isolated. It does not prove privilege was preserved, which remains a court's determination. Do not trust any claim that a tool guarantees privilege.

Factors to work through with your own counsel

None of this has to be reinvented matter by matter. The same handful of questions comes up each time, and a firm that answers them before client material reaches a tool is running the confidentiality analysis by habit rather than by luck. These questions map to the factors laid out above, and they are meant to structure a conversation with a licensed attorney in your jurisdiction, not to substitute for one.

This article is general information from a founder who builds verification tools for law firms, not legal advice, and not an opinion on any specific matter. ABA opinions are non-binding guidance that states adopt differently, the confidentiality duty is distinct from privilege, and privilege waiver is decided case by case. Use the questions below as a starting point, then take the actual decision to your own counsel and your jurisdiction's rules.

• Does the input include information relating to the representation, and how sensitive is it ^[1]?
• Is the tool self-learning, or does it carry a contractual no-training and zero-retention commitment ^[1]?
• Has the client given informed, specific consent covering this use, or only signed boilerplate ^[1]?
• Does your jurisdiction's own guidance add requirements around disclosure, billing, or oversight ^[3][4]?
• Have you run the privilege question as a separate analysis, recognizing it is unsettled and decided by courts?
• Can the firm later show, not just assert, that consent was captured and sensitive material was isolated?

This is informational, not legal advice. Because this is an ethics question that varies by jurisdiction and by facts, consult a licensed attorney in your jurisdiction before relying on any AI workflow.