Supervising AI: A Partner's Duty Under Rules 5.1 and 5.3

When a junior lawyer or paralegal uses a generative AI tool to draft a brief, the work that results is supervised work. The partner who signs it, or who runs the firm where it is produced, carries responsibility for it under the same rules that have always governed delegation. The technology is new. The supervisory duty is not.

This piece reads that duty through three sources every managing partner should know by name: Model Rule 5.1, Model Rule 5.3, and ABA Formal Opinion 512. It then sets out what reasonable supervisory measures look like in a firm of three to thirty lawyers, where there is no general counsel, no innovation committee, and no margin for a sanctions order with the firm's name in the caption.

Two rules, two halves of the same obligation

Model Rule 5.1 is titled "Responsibilities of Partners, Managers, and Supervisory Lawyers." It does three things. Subsection (a) requires a partner, and any lawyer with comparable managerial authority, to make reasonable efforts to ensure the firm has measures in effect giving reasonable assurance that all lawyers conform to the rules. Subsection (b) puts a parallel duty on any lawyer with direct supervisory authority over another lawyer. Subsection (c) makes a lawyer responsible for another lawyer's violation if the lawyer ordered or ratified the conduct, or, being a partner, manager, or direct supervisor, knew of it at a time when its consequences could be avoided and failed to act.

Model Rule 5.3 carries the same architecture over to non-lawyers. Its title is worth pausing on. In 2012 the ABA changed the heading from "Nonlawyer Assistants" to "Responsibilities Regarding Nonlawyer Assistance." The shift from a noun describing people to a word describing services was deliberate. It extends the rule beyond paralegals and secretaries to outsourced and technological assistance the firm relies on. That is the textual hook for treating AI output as something a lawyer must supervise, and it is why the phrase Rule 5.3 AI nonlawyer assistance has become shorthand for the firm's duty over the tools it deploys. The subsections mirror 5.1: managerial lawyers must put measures in place, supervising lawyers must make reasonable efforts, and a lawyer is responsible for a non-lawyer's conduct that would violate the rules if the lawyer ordered, ratified, or failed to mitigate it.

Read together, the rules describe a chain. The firm must have a system. The supervisor must work the system. And when something goes wrong that a reasonable supervisor would have caught, the failure to supervise is itself the violation, independent of whoever pressed the keys.

What ABA Opinion 512 actually says

The ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, "Generative Artificial Intelligence Tools," on July 29, 2024. It is the first comprehensive national ethics guidance on lawyers' use of generative AI, and it speaks directly to supervision.

On the managerial side, the opinion states that lawyers with managerial authority should establish clear policies on the firm's use of generative AI and provide for training of both lawyers and non-lawyers in the firm. The opinion describes the substance such training should cover: the basics of how the technology works, the capabilities and limitations of the tools, the ethical issues their use raises, and best practices for secure handling of data, privacy, and confidentiality. The opinion frames this as an ongoing obligation rather than a one-time orientation, reflecting that the tools change quickly.

The opinion also reaches outside the firm's walls. It notes that supervisory responsibility extends to situations where lawyers rely on others outside the firm in connection with a representation, which can include cloud platforms that store client data and third-party service providers. In other words, choosing a vendor is itself a supervised act.

Two practical points from the opinion deserve emphasis for smaller firms. First, the opinion treats independent verification of generative AI output as central. A lawyer cannot rely on the tool's self-assessment of its own accuracy. Second, the opinion makes clear that the duty of competence under Rule 1.1 is engaged: a lawyer must have a reasonable understanding of the capabilities and limitations of a tool used in a representation. Supervision under 5.1 and 5.3 sits on top of that competence baseline.

The cases show how this fails in practice

The risk here is not hypothetical, and the published sanctions decisions are the clearest teaching tool available.

The original cautionary tale is Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023). Judge P. Kevin Castel sanctioned two lawyers and their firm, Levidow, Levidow & Oberman, after a brief cited cases that did not exist, generated by ChatGPT. The court imposed a $5,000 penalty and found the lawyers had acted with subjective bad faith, in part because they failed to read the cited cases or take any step to confirm the authorities were real. The lesson is not "do not use AI." It is that a citation no one read is a citation no one supervised.

For supervision specifically, the more instructive decision is Wadsworth v. Walmart, Inc. (D. Wyo. 2025). Lawyers from Morgan & Morgan filed motions in limine citing nine cases, eight of which did not exist, produced using an in-house AI tool. Judge Kelly Rankin sanctioned not only the drafting attorney, Rudwin Ayala, who was fined $3,000 and had his pro hac vice admission revoked, but also the supervising attorney and local counsel, who were each fined $1,000. The supervisors had not drafted the motions and, by the record, had not reviewed them before filing, yet they had e-signed the documents. That is the supervisory point in stark form: a signature on a delegated filing is an assertion that the work was checked, and the duty does not transfer to the most junior person in the chain.

These outcomes track the structure of Rules 5.1 and 5.3. The drafter bears responsibility for the work. The supervisor bears a separate responsibility for the failure to supervise. "The AI wrote it" is not a defense, because the rule was never about who typed the words.

What reasonable supervisory measures look like

Neither the rules nor Opinion 512 demand that a partner re-do every associate's research. The standard is reasonable efforts and reasonable assurance, not perfection. For a small or mid-sized firm, reasonable measures are concrete and inexpensive. The following is a working checklist, not legal advice, and it should be adapted to your jurisdiction's adopted rules and any local court standing orders.

• Adopt a written AI policy. State which tools are approved, what client information may and may not be entered into them, and who to ask when in doubt. A short, clear policy beats a long one no one reads.
• Require human verification of every cited authority. The rule is simple: no citation, quotation, or proposition of law leaves the firm unless a lawyer has pulled the source and confirmed it says what the draft claims. Make this a documented step, not a cultural assumption.
• Define a confidentiality line for inputs. Decide whether client identifiers, privileged facts, and confidential documents may be entered into a given tool, and understand the vendor's data handling and training practices before you decide. Treat tool selection as a supervised vendor decision under Rule 5.3.
• Build training that repeats. Cover how the tools generate plausible but false output, where they fail, and the firm's verification and confidentiality rules. Refresh it as tools change, and include non-lawyer staff, who Opinion 512 expressly contemplates.
• Match review to risk. A first-pass internal memo and a dispositive motion do not warrant the same scrutiny. Calibrate supervisory review to the stakes, the audience, and the experience of the person doing the work.
• Decide on client disclosure deliberately. Whether and how you tell clients about AI use in their matters is a judgment call that intersects with communication and, in some engagements, billing. Make it a considered policy, not an afterthought.
• Keep a light record. A note that the policy exists, that training occurred, and that verification was performed is what turns "we are careful" into demonstrable reasonable efforts if anyone ever asks.

The connection to the rules is direct. Adopting a policy and training people is the firm-level measure that Rule 5.1(a) and Rule 5.3(a) require of managerial lawyers. Verifying authorities before a filing goes out is the supervisory effort that Rule 5.1(b) and Rule 5.3(b) require of the lawyer running the matter. Doing both is also the practical difference between the firm in Wadsworth that filed unreviewed motions and a firm that catches the error in-house.

Supervising AI without slowing the firm down

The reasonable reading of all of this is not caution for its own sake. Generative AI can genuinely help a small firm draft faster and research more broadly, and Opinion 512 does not discourage its use. The point of supervising AI in a law firm is to capture that benefit while keeping the firm's name off a sanctions docket and its client relationships intact.

The measures that accomplish that are modest. A one-page policy, a verification habit your lawyers actually follow, training that is repeated rather than performed once, and a clear-eyed view of which tools touch confidential data. None of this requires a dedicated budget or a technology officer. It requires a partner to decide that the firm has a system and to make sure people use it.

That decision is exactly what Rules 5.1 and 5.3 have always asked of the people in charge. The tools changed. The job did not.

This is general information for lawyers and law-firm leaders, not legal advice, and it does not create an attorney-client relationship. The authorities are cited so you can read them yourself.