The Lawyer's Duty to Supervise AI Tools: Model Rules 5.1, 5.3, and State Guidance

A managing lawyer looking at the latest sanctions order does not need another speech about whether generative AI is exciting or dangerous. The harder question is more immediate: if someone on the team used an AI tool before a filing, memo, contract review, or client advice left the office, what exactly was the supervising lawyer supposed to have done first?

The short answer is that the lawyer duty to supervise AI tools already sits inside familiar ethics rules. Model Rule 5.1 requires partners and lawyers with managerial authority to make reasonable efforts to ensure that the firm has measures giving reasonable assurance that all lawyers conform to the Rules of Professional Conduct, and it imposes separate duties on lawyers with direct supervisory authority over another lawyer.[1] Model Rule 5.3 does the same work for nonlawyer assistance, including the duty to ensure that nonlawyer conduct is compatible with a lawyer’s professional obligations.[2] ABA Formal Opinion 512 then makes the AI application explicit: lawyers using generative AI must consider policies, training, and verification as part of their supervisory responsibilities.[3]

That answer matters because AI use has already moved from novelty to office infrastructure. A 2025 Thomson Reuters survey reported that 79% of North American legal professionals use AI.[4] Adoption, though, is not the same as supervision. A firm can have broad use and still have no answer when a court asks who checked the cases, who approved the prompt workflow, who trained the paralegal, or who decided that a vendor’s confidentiality terms were acceptable.

The rule foundation was already there

The important move in Rule 5.3 happened well before the current wave of generative AI products. In 2012, the ABA changed the title of Model Rule 5.3 from “Responsibilities Regarding Nonlawyer Assistants” to “Responsibilities Regarding Nonlawyer Assistance.”[2] That change is easy to pass over. It should not be.

The old title pointed attention toward a person: the paralegal, investigator, clerk, contract reviewer, vendor employee, or outside service provider. The revised title points toward the function: assistance in legal work performed by something or someone that is not a lawyer. That is why AI did not need a bespoke ethics category before it could become supervisable. If a tool is supplying research leads, draft language, summaries, deposition outlines, issue lists, privilege calls, or contract analysis that a lawyer may use in representing a client, it is participating in legal work. The lawyer remains responsible for whether that work product is competent, confidential, accurate, and appropriately communicated.

Rule 5.1 supplies the managerial side of the same obligation. It is not limited to the lawyer who personally typed a prompt. Partners and lawyers with comparable managerial authority must make reasonable efforts to establish internal measures that give reasonable assurance of compliance with the ethics rules.[1] A supervising lawyer must also make reasonable efforts to ensure that a subordinate lawyer conforms to those rules.[1] That is the part firms sometimes miss when they frame AI use as a private productivity choice by individual lawyers.

A private experiment becomes an institutional problem the moment the output enters client work. The associate who pastes a fake citation into a brief may draw the court’s first anger. The paralegal who relies on a summary may be the easiest person to identify. But Rules 5.1 and 5.3 ask a prior question: what reasonable system did the supervising lawyer create before those people were put in that position?

What ABA Formal Opinion 512 adds

ABA Formal Opinion 512, issued in July 2024, does not treat generative AI as an ethics-free shortcut or as forbidden machinery. Its value is more practical: it translates familiar duties into management tasks. For supervision, the opinion identifies the need for lawyers with managerial and supervisory authority to address generative AI through clear policies, training of lawyers and nonlawyers, and verification of AI-generated work product.[3]

The word “verification” deserves more weight than it usually receives in vendor materials. It is not a ceremonial glance at a polished answer. In litigation, it may mean pulling every cited authority, confirming quotations and pin cites, checking whether a case is still good law, and comparing the AI-generated proposition to the actual holding. In transactional work, it may mean testing a clause summary against the full agreement, confirming defined terms, and reviewing whether omitted language changes the risk allocation. In investigations or due diligence, it may mean preserving the difference between a document-grounded extraction and a model-generated inference.

Opinion 512 also keeps responsibility with lawyers who supervise people, not just with people who operate tools. A firm that allows junior lawyers, paralegals, contract attorneys, or legal operations staff to use generative AI without guidance has not delegated risk; it has created it. The downstream user may make the visible mistake, but the upstream failure is often the absence of a rule, a training record, a review checkpoint, or a named lawyer accountable for the workflow.

State bars are translating the same baseline into local expectations

State guidance is developing unevenly, as state guidance always does. The useful point is not that every jurisdiction has used the same words. It is that the leading opinions are converging around the same supervisory instincts: review the output, train the people using the tools, protect client confidences, and do not let convenience substitute for professional judgment.

Florida makes the supervision point teachable

Florida Bar Ethics Opinion 24-1 treats AI output as subject to Rule 4-5.3 supervision and says lawyers must review and verify the accuracy and sufficiency of work product produced with generative AI.[5] The opinion’s most useful teaching device is its analogy to a junior associate. A partner may assign work to a junior lawyer, but the partner cannot file the assignment unchecked because it looks plausible. The same principle applies when the first draft comes from a machine.

That analogy works because it is concrete. It does not require lawyers to become software engineers before they can supervise AI-assisted work. It does require them to understand enough about the tool and the task to know what must be checked. A lawyer who would never let a first-year associate submit an uncited research memo as a final brief should not let a chatbot do it under the softer label of innovation.

Oregon, North Carolina, and Texas emphasize existing duties

Oregon Formal Opinion 2025-205 addresses lawyers’ use of artificial intelligence under existing professional conduct rules, including the need to preserve confidentiality and exercise appropriate professional judgment.[6] North Carolina 2024 Formal Ethics Opinion 1 likewise frames AI use through existing duties rather than a new technology exception.[7] Texas Opinion 705, issued through the Texas Center for Legal Ethics, similarly treats AI as a tool that must be used consistently with established obligations, including competence, confidentiality, and supervision.[8]

For a multistate practice, the lesson is not that one national memo can be pasted into every office manual and forgotten. The baseline may be familiar, but implementation is jurisdictional. A legal department with lawyers in Florida, Texas, Oregon, and North Carolina should expect the same general questions to produce different local phrasing, different disclosure instincts, and different risk tolerances. Someone has to monitor that.

California is important, but still proposed

California’s proposed amendments deserve close attention because they would state the review obligation in especially direct terms. As reported in May 2026, the proposal would require a lawyer using generative AI to “independently review, verify, and exercise professional judgment regarding any output generated by the technology that is used in connection with representing a client.” Public comment closed in May 2026.[9]

That is not the same as current final law. The proposal may be modified before any adoption by the California Supreme Court. Still, it is a serious signal. If adopted in substantially similar form, California would move from general supervision language to an explicit independent-review requirement for AI outputs used in client representation. Firms with California lawyers should not treat that as a remote drafting exercise. They should be asking now whether their AI workflows could satisfy that standard if it became enforceable.

Courts are punishing failures after the handoff

The sanctions cases are often described as hallucination stories. That is too narrow. A hallucinated case is the visible defect. The professional failure is the handoff: AI output moved into a court filing without adequate human verification.

The cases also should not be flattened into a single escalation chart. They arise in different procedural settings, under different authorities, and with different consequences. Still, the pattern is hard to ignore. Norton Rose Fulbright’s March 2026 update, citing Damien Charlotin’s database, reported 1,148 documented hallucination cases by U.S. lawyers as of early 2026; that is a tracker estimate, not an official court count.[10]

The Fifth Circuit’s language in Fivehouse is especially important because it changes the tone of the enforcement conversation. The court warned that courts “must move beyond admonitions and reprimands into more punitive sanctions.”[10] That sentence is not a new ethics rule. It is worse for anyone relying on informal habits: it is a judicial patience report.

A supervisor reading that line should hear a practical demand. If a filing contains AI-generated errors, the court may ask what the filing lawyer did. A client, insurer, disciplinary authority, or firm management committee may ask what the supervising lawyer required before the filing lawyer ever reached that point. The answer cannot be “we told people to be careful.”

What a defensible supervision system should be able to show

A defensible system does not have to be ornate. It does have to be describable. If the firm’s policy exists only as hallway advice, a partner’s preference, or a vendor slide saying “human in the loop,” it will be difficult to prove that anyone exercised the reasonable efforts contemplated by Rules 5.1 and 5.3.

• A written AI-use policy that identifies approved tools, prohibited tools, permitted use cases, confidentiality restrictions, approval procedures, and consequences for noncompliance.
• Training for lawyers and nonlawyers that covers tool limits, hallucination risk, confidentiality, privilege, client disclosure issues, and the difference between drafting assistance and verified legal work.
• A verification requirement for AI outputs used in client work, including checking legal authorities, quotations, record citations, factual summaries, calculations, contract provisions, and jurisdiction-specific law.
• A review chain that identifies who must approve AI-assisted work before it goes to a client, court, regulator, counterparty, board, or business unit.
• Documentation or audit habits where appropriate, especially for litigation filings, high-risk advice, client-confidential uploads, and workflows performed by nonlawyers or junior lawyers.
• A monitoring process for state guidance, court orders, local rules, and proposed rule changes, including California’s pending amendments if relevant to the practice.

The right level of documentation will vary. A routine internal brainstorming prompt does not need the same record as a dispositive motion, opinion letter, privilege review protocol, or regulated-industry client advice. But the judgment should be made in advance, not reconstructed after an error. Risk committees know the difference between a calibrated system and a scramble.

Verification should also be tied to the output’s use. If an AI tool generates possible research paths, the lawyer may need to confirm the authorities before relying on them. If it summarizes a document set, the reviewer may need to test the summary against source documents. If it drafts language for a client deliverable, the responsible lawyer must decide whether the language is legally sufficient, factually accurate, and appropriate for the client’s objective. The tool can accelerate the path to a draft. It cannot assume the lawyer’s duty of judgment.

Supervision of nonlawyers needs particular care. A paralegal, legal assistant, investigator, or contract reviewer may use an approved AI tool exactly as trained and still produce output that requires lawyer review before use. That is not a failure by the nonlawyer. It is how Rule 5.3 works. The lawyer must make reasonable efforts to ensure that the assistance is compatible with professional obligations, including confidentiality, competence, candor, and communication duties.[2]

Supervision of subordinate lawyers is only slightly different. Junior lawyers can and should learn efficient tools, but they also need to know when AI assistance ends and legal responsibility begins. A supervising lawyer who wants associates to use generative AI for first drafts should say what must be independently checked, what sources count as authoritative, whether prompts or outputs should be saved, when client information may be entered, and when a partner must be told that AI materially assisted the work.

The ethical issue in 2026 is no longer whether lawyers may experiment with AI. Many already do. The issue is whether lawyers with managerial and supervisory authority can show that they exercised professional judgment over the people and tools producing legal work before the mistake reached a client, a court, or a regulator.

References

1. Rule 5.1: Responsibilities of Partners, Managers, and Supervisory Lawyers, American Bar Association, link
2. Rule 5.3: Responsibilities Regarding Nonlawyer Assistance, American Bar Association, link
3. Formal Opinion 512: Generative Artificial Intelligence Tools, American Bar Association, July 29, 2024, link
4. 2025 Generative AI in Professional Services Report, Thomson Reuters, 2025, link
5. Ethics Opinion 24-1, The Florida Bar, January 19, 2024, link
6. Formal Opinion 2025-205, Oregon State Bar, 2025, link
7. 2024 Formal Ethics Opinion 1, North Carolina State Bar, 2024, link
8. Opinion 705, Texas Center for Legal Ethics, link
9. California Bar Advances Proposed Rules on Lawyers’ Use of Generative AI, LawNext, May 2026, link
10. AI hallucinations and legal sanctions: courts move beyond warnings, Norton Rose Fulbright, March 2026, link