How to Supervise AI Like You’d Supervise a Junior Associate

The surveys don’t measure the same thing, but they point in the same direction.

The Clio 2025 Legal Trends Report found that 79% of legal professionals use AI. The Wolters Kluwer 2026 Future Ready Lawyer Survey puts that figure at 92%. The 8am 2026 Legal Industry Report, surveying over 1,300 legal professionals, found 69% using generative AI for work, more than doubled from 31% the prior year.

Now the other side. Only 9% of firms have a written, actively enforced AI policy (8am 2026). 43% have no policy and no plans to create one. 54% provide no AI training and plan none.

The gap between individual adoption and institutional oversight is where the professional responsibility problem lives. Rule 5.1(a) requires partners and managing lawyers to establish “measures giving reasonable assurance” that everyone at the firm follows the Rules of Professional Conduct. When the vast majority of your team uses AI but your firm has no written guidance on how, you’ve got a 5.1(a) problem whether you know it or not.

And the underground AI issue makes things worse. Axiom Law’s 2024 survey of 300 in-house counsel found that 83% use AI tools their company didn’t provide, and 81% acknowledged using tools that haven’t been formally approved. Tomasz Zalewski, writing in the Wolters Kluwer Future Ready Lawyer report, put it directly: “If they do not get secure tools, they will have ‘shadow AI.’ Bans and restrictions do not work.”

Here’s what makes this specifically dangerous for smaller firms. A Stanford Cyberlaw study from October 2025 analyzed 114 attorney cases and found that 90% of cases involving AI-tainted filings came from solo or small practices. Solo practitioners and small firms aren’t just using AI without supervision. They’re the ones getting caught. For more on the financial and professional consequences of these failures, see our analysis of AI malpractice risk for lawyers.

Many firms handle AI responsibly through existing quality control, partner review of filings, and cite-checking workflows. A written policy is not the only way to satisfy Rule 5.1(a), but it is the easiest to defend when a disciplinary committee asks what measures you had in place. The question isn’t whether your firm has had a problem. It’s whether your current process would hold up to a disciplinary committee’s inquiry into what “reasonable efforts” you made.

The idea of treating AI like a junior team member didn’t start as a blog post headline. It comes from the ethics opinions themselves.

The D.C. Bar, in Ethics Opinion 388 (2024), offered the most memorable framing: AI is “an omniscient, eager-to-please intern who sometimes lies to you.” A British Institute of International and Comparative Law report advises lawyers to “treat AI tools like young associates,” noting they “might be a fantastic asset, but lawyers need to invest, supervise, verify, have a little bit of scepticism.” DLA Piper describes AI as a “subordinate” requiring “direction and supervision.” And Professor Michael Murray’s scholarship frames AI as a “staff attorney” that can be valuable “if used in appropriate situations and under constant supervision by human lawyers.”

The analogy works because it maps onto instincts you already have. You know what it means to supervise a first-year. You review work product before it goes out. You check the reasoning, not just the conclusion. You verify citations. You ask follow-up questions when something feels off. And you make the final call on strategy and judgment. None of that changes when the work product comes from software instead of a person.

But the analogy has limits, and those limits matter. A junior associate learns from correction; the next memo improves because of your feedback. Most AI tools don’t work that way. They won’t internalize your firm’s standards over time or develop professional judgment through mentorship. AI also operates at a scale that makes full review harder. A first-year might draft three memos a week. An AI tool can produce thirty in a day.

And that’s precisely why the analogy is useful in the direction that matters: AI requires more supervision than a junior associate, not less. It can’t learn, can’t self-correct, and can’t flag its own uncertainty. When a first-year doesn’t know an answer, they ask you. When AI doesn’t know an answer, it makes one up. Stanford HAI research found that general-purpose chatbots hallucinated 58% to 82% of the time on legal queries.

You’d never let a first-year file a brief you hadn’t read. Same standard.

In early 2023, attorney Steven Schwartz cited six fabricated judicial opinions in a brief filed in Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023). He used ChatGPT for legal research and never checked whether the cases existed. The attorney of record, Peter LoDuca, signed the brief without reading a single cited case. Judge P. Kevin Castel imposed a $5,000 penalty. The fine was modest. The reputational damage was not.

Then the delegation chains got longer. In Mezu v. Mezu, No. 361, Sept. Term 2025 (Md. Ct. Spec. App.), an attorney admitted he never read the cases cited in his brief. He had relied on his law clerk, a nonlawyer, who also never read them. The law clerk used ChatGPT to search for relevant cases, received a generated list that included nonexistent citations, then searched Google, VLex, and other platforms to extract what looked like proper citations. Attorney to nonlawyer to AI to nobody verifying anything. The court referred the attorney to the Attorney Grievance Commission. This is the cascading failure Rule 5.3 was written to prevent.

Having a policy doesn’t fix this if the people enforcing it break it themselves. In Johnson v. Dunn, No. 2:21-cv-1701 (N.D. Ala. July 2025), Butler Snow, a large, well-regarded firm, had established AI policies including practice group leader approval requirements. But an attorney on the team inserted ChatGPT-generated citations into a motion without verification. The court publicly reprimanded the attorneys, disqualified them from the case, and referred them to the Alabama State Bar, declaring that fines and public embarrassment were not effective deterrents and that substantially greater accountability was needed. For a breakdown of the 300-plus judicial AI standing orders now in effect, see our guide to AI disclosure rules.

Proprietary tools don’t solve the problem either. In Wadsworth v. Walmart Inc., 348 F.R.D. 489 (D. Wyo. 2025), Morgan & Morgan attorney Rudwin Ayala used the firm’s own AI tool, “MX2.law.” Eight of nine cited cases did not exist. The supervising attorney directed the work but never reviewed it. The court sanctioned three attorneys a total of $5,000, with Ayala receiving the largest share at $3,000 and having his pro hac vice admission revoked.

These cases are sometimes dismissed as “bad lawyer” problems, not AI problems. That’s partly right, and that’s the point. Every sanctioned attorney failed at something basic: reading their own filing before submitting it. But AI makes this failure easier to commit, harder to catch, and faster to scale. Damien Charlotin’s global database now tracks over 1,200 cases. Among the millions of filings submitted annually, that is a small fraction, but the trajectory is accelerating and the consequences are intensifying. The trajectory runs from warnings (2023) to fines to disqualification, bar referrals, and license suspension. Courts have made clear that the standard isn’t changing. The consequences are.

The supervision duty runs in two directions. Misuse AI, and you face sanctions. Ignore AI, and you may face a different kind of professional exposure.

Dean Andrew Perlman of Suffolk University Law School has made the most provocative version of this argument: “The duty of competence may eventually require lawyers’ use of generative AI. The technology is likely to become so important to the delivery of legal services that lawyers who fail to use it will be considered as incompetent as lawyers today who do not know how to use computers, email, or online legal research tools.”

That’s not settled law. It’s an argument from one of the profession’s most credentialed ethics scholars, and it hasn’t been adopted by any state bar. But the direction is clear enough to take seriously.

Comment [8] to Rule 1.1, amended in 2012, says competent practice includes understanding “the benefits and risks associated with relevant technology.” Forty states, the District of Columbia, and Puerto Rico have adopted that language. The ABA’s 2025 Task Force report concluded that AI has moved “from experiment to infrastructure” for the legal profession.

And the insurance market is already pricing the risk. Some malpractice carriers are adding AI-specific sublimits or exclusions to professional liability policies. Berkley Insurance introduced what it calls an “Absolute” AI exclusion, covering claims related to “any actual or alleged use, deployment, or development of Artificial Intelligence.” If your policy has an AI exclusion and you’re using AI without documented supervision protocols, you may be uninsured for the exact scenario these rules are designed to prevent.

The duty isn’t to avoid AI. It’s to supervise it. The same way you’d supervise a junior associate who’s smart, fast, confident, and occasionally fabricates the law.