Do Lawyers Have to Disclose AI Use? To Whom, When and in What Words

In Clio’s 2026 UK and Ireland survey, 81% of firms said they disclose their AI use to clients at least occasionally. Seven per cent of clients recalled being told. The ACC and Everlaw found that 59% of in-house legal teams do not know whether their outside firms use generative AI on their matters at all. Disclosure is happening on paper and not in anyone’s memory.

That is because “do lawyers have to disclose AI use” is three questions wearing one coat. The client, the court and the client-as-buyer each have a different source of duty, a different trigger and a different form of words. Conflate them and you either over-disclose (a boilerplate paragraph nobody reads) or under-disclose (client data in a self-learning tool, no consent). Here is the separation, and the sentences to use.

Three audiences, three different rules

The ethics and regulation hub and the jurisdiction guide have the opinions; this page is about what to say.

Clients: when you must disclose AI use and obtain consent

ABA Formal Opinion 512 (29 July 2024) draws the line at self-learning tools. Where the tool learns from inputs, “a client’s informed consent is required prior to inputting information relating to the representation.” And the sentence that kills most engagement-letter paragraphs:

“Merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient.” — ABA Formal Opinion 512, Generative Artificial Intelligence Tools (29 July 2024)

Informed consent needs four things: the lawyer’s best judgement about why the 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 tool’s benefits. Two further triggers: disclosure is mandatory when the client asks, and when AI use is relevant to the basis or reasonableness of the fee.

The state opinions agree on the trigger. Florida 24-1 recommends informed consent before any third-party generative AI use that “would involve the disclosure of any confidential information”. North Carolina’s 2024 FEO 1 adds a second trigger: delegating “substantive tasks in furtherance of the representation” to an AI tool requires “advanced informed consent”.

Clients: when it is not

Routine use in a closed tool needs no announcement. North Carolina: “Generally, a lawyer need not inform her client that she is using an AI tool to complete ordinary tasks, such as conducting legal research or generic case/practice management.” Virginia: “There is no per se requirement to inform a client about the use of generative AI in their matter.” New Jersey requires disclosure only “if a client asks … or if the client cannot make an informed decision about the representation without knowing”. Germany’s BRAK finds no professional duty to inform, while recommending “im Zweifel eine vertragliche Regelung mit den Mandanten”.

Pennsylvania’s Joint Formal Opinion 2024-200 says lawyers “should be transparent with clients, colleagues, and the courts about the use of AI tools”, and the CCBE’s 2025 guide asks you to be transparent “if it can reasonably be assumed that an informed client would object, make conditions, or otherwise have reservations”. A Copilot summary inside your own Microsoft tenant, with enterprise data protection on, fails none of these tests; a witness statement drafted in a consumer chatbot fails all of them.

Courts: standing orders and certification language

Judge Brantley Starr’s order (N.D. Tex., 30 May 2023) requires a certificate that either no portion of the filing was AI-drafted or that a human checked it against “print reporters or traditional legal databases”. The Legal AI Governance tracker counts 113 active orders, 82 of them requiring verification or disclosure; the Fifth Circuit and the Illinois Supreme Court rejected blanket rules, the Fifth Circuit adding that “‘I used AI’ will not be an excuse for an otherwise sanctionable offense.”

In Lnu v. Blanche (9th Cir., 3 June 2026) two lawyers, one of whom had denied AI use three times at argument, were fined $2,500 each, suspended from Ninth Circuit practice for six months, and ordered, with every attorney at their firm, to include sworn AI disclosures in every filing for two years. Canada’s Federal Court expects a first-paragraph declaration. The order types, the international rules and a model certificate are in the court standing orders guide.

Courts: the UKUT statement-of-truth formula

England and Wales does not ask whether you used AI; it asks you to swear to the authorities. After a solicitor filed a case that Google’s “AI Mode” had invented, the Upper Tribunal in UK and Munir v SSHD [2026] UKUT 81 (IAC) changed the judicial-review claim form. The legal representative must now confirm by statement of truth that every authority cited “(a) exists; (b) may be located using the citation provided; and © supports the proposition of law for which it is cited.”

That formula is a better disclosure than “AI was used”: it tells the court the one thing it cares about.

Opponents and outside counsel guidelines

No US rule requires you to tell opposing counsel that you used AI, unless a standing order or, as in Canada, a court notice says otherwise. The duty that does run to the other side is the reverse one: in Noland v. Land of the Free (Cal. Ct. App., September 2025) the respondents were denied their fees because they “did not alert the court to the fabricated citations” in the appellant’s brief.

Clients as buyers are different: outside counsel guidelines are absorbing AI clauses fast; Layer3Labs’ summary of the standard provisions lists disclosure and prior written consent, approved and prohibited tools (many OCGs ban free public chatbots for any client work), no client data into public models, billing for actual time with no charge for learning tools, and mandatory human review. Poppy Legal’s model billing clauses add the detail: “Company data may not be used to train a provider’s or third party’s models”; no reconstructed “equivalent time”; and an invoice notation such as “AI-assisted; attorney reviewed”.

The ACC’s 2025 finding that 59% of in-house teams do not know whether their firms use generative AI is an opportunity. The firm that volunteers a clear answer before the OCG demands one keeps the client; the OCG guide covers the negotiation.

Marketing chatbots: the public-facing rule

Rule 7.1 applies to AI-written marketing exactly as it applies to a brochure: “The ‘AI drafted it’ defense does not exist,” as one compliance overview puts it. Florida 24-1 warns against a chatbot that may “fail to immediately identify itself as a chatbot”, and several states (New York, California, Colorado, Utah and New Jersey, according to the same overview) have consumer statutes requiring bots to say they are not human. The advertising rules guide has the state detail. In the EU, Article 50 of the AI Act applies from 2 August 2026: a chatbot must tell users they are dealing with a machine, while text that has undergone human review needs no label under Art. 50(4).

Sample engagement-letter clause

Written to the four elements of ABA 512; delete anything untrue of your practice.

Draft an engagement-letter clause on our use of AI tools for a [law firm] in [jurisdiction] that satisfies the informed-consent standard in ABA Formal Opinion 512 rather than boilerplate. Cover: the tools we use ([e.g. ChatGPT Business, Claude Team, Copilot with enterprise data protection]) and their no-training and retention terms; what client information may be processed; the specific risks (error, confidentiality, retention, provider access) and the benefits; our human-review commitment; the client's right to object; and billing (actual time only, no charge for learning tools). Plain English, under 250 words.

Sample filing certification

A composite of Judge Starr’s two branches, the Lnu order and the Upper Tribunal formula; useful as a discipline even where no court requires it.

From <document>...</document>, list every case, statute, rule and secondary source cited in a table: Citation as written | Proposition (quote my sentence) | Pinpoint | Quotation? (Y/N) | Verified by | Database | Date. Leave the last three columns blank; I will complete them by hand. Do not tell me whether any citation exists or is accurate.

Sample OCG response

When a client’s guidelines ask “Do you use AI on our matters, and how?”, the answer that works is short, specific and volunteers the controls. If an AI-assisted filing later goes wrong, the fake-citation playbook applies to clients as much as courts.

Draft our response to the AI section of [client]'s outside counsel guidelines <ocg>...</ocg>. Facts, which you must not embellish: we use [tools and tiers]; none trains on inputs, retention is [period]; we do not use consumer chatbots for client work; every AI-assisted output is reviewed by the responsible lawyer; we bill actual time only and note "AI-assisted; attorney reviewed" on relevant entries. For each OCG requirement, state Comply / Comply with clarification / Cannot comply, with one sentence each; where we do not do something, say so and propose an alternative. Under 400 words.

The ABA Formal Opinion 512 explainer has the consent and fee rules, the client communication guide covers explaining AI use in plain English, and the prompt library has the full set. Knowing which tier of ChatGPT, Claude, Gemini or Copilot you are on is a five-minute settings exercise that every disclosure clause depends on; in AI Lab for Lawyers we do it live.