AI is the new associate — it drafts the brief, builds the cites, and never once suspects it's wrong. In our latest blog, "Meet the New Boss," Leon Silver lays out what real AI supervision looks like, and why the question a court asks after a fake citation is never what the AI did, but what the lawyer did to catch it. In the next post, I'll show you what happens when you comply with an AI disclosure order to the absolute letter — and why the result often is an indictment of the rule, not the practice.
What the Rules Already Require
The ethical framework governing AI supervision is not waiting for bar associations to issue new rules. The existing rules are sufficient, and courts are applying them.
Arizona’s Ethical Rule 1.1, like the ABA Model Rule on competence, requires that lawyers provide competent representation — including the legal knowledge, skill, thoroughness, and preparation reasonably necessary for the representation. The comments to that rule have been updated to make clear that competence includes understanding the benefits and risks of relevant technology. A lawyer who uses AI tools without understanding how they work, what they get wrong, and how to verify their output is not meeting that standard, regardless of the quality of the output on any given day.
ER 5.3 requires that lawyers with supervisory authority over non-lawyer assistants ensure that their work is compatible with the lawyer’s professional obligations. The argument that this rule extends to AI is gaining traction among bar associations, and it is persuasive: if you are using AI to do work that a paralegal or junior associate would otherwise do, the supervision obligation does not disappear because the assistant is a machine. If anything, it becomes more demanding, because the machine does not know when it is wrong and will not tell you. (Practice tip: If you aren’t telling AI when it’s wrong, which is more often than you’d expect, you’re not acting like a boss at all.)
As we discussed in our post on Heppner and the privilege implications of AI use, courts are already examining the specific tools lawyers use and the specific steps they took — or failed to take — to verify AI output. The supervision failure that leads to sanctions is not treated as a technology problem. It is treated as a professional failure.
What Most Firms Call Supervision
In practice, AI supervision in most firms means reading the output. A lawyer or a senior associate reviews the AI-generated draft, checks it for obvious errors, adjusts the tone, and sends it along. Citations are assumed to be correct because they look correct. Factual assertions are assumed to be accurate because they are stated with confidence. The review is real — it is just not the right review.
Reading for quality is not the same as verifying for accuracy. AI systems are very good at producing output that reads well. They are not reliably accurate, and their errors are not random — they are systematically plausible. A hallucinated case citation does not look like gibberish. It looks like a real case, from the right court, with the right format, holding something the AI calculated would support your argument. Catching it requires something the read-through does not provide: verification against the primary source.
The same problem applies to factual assertions, regulatory citations, procedural rules, and any other content the AI has generated from its training data rather than from documents you provided. The AI does not distinguish between what it knows and what it has confabulated. Both come out in the same confident register.
What Real Supervision Looks Like
My own practice on citations is a useful illustration of what the supervision obligation actually requires, and the mechanic is not complicated. We draft without hyperlinks — every citation goes into the document as plain text. Before any work product leaves the office, the supervising lawyer pulls each cite directly on Westlaw or Lexis, confirms the case exists, confirms it stands for what it is cited for, and confirms it is accurately and correctly cited. Only then does the lawyer copy the live hyperlink from Westlaw or Lexis into the document. The hyperlink is not decoration — it is proof of work. A linked citation is one a human being personally retrieved, read, and verified; an unlinked one never made it to the desk. The byproduct is a brief the court can navigate with a click, which judges appreciate and some now require — so the same step that protects us against a fake citation also hands the court something it wants. If a court ever questions a citation in our work product, we can show exactly when we pulled it and what we found.
That is not an elaborate process. It takes time, but it is time that would have been spent anyway in a properly supervised traditional workflow. What it is not is optional. A citation that has not been verified is a liability waiting to be triggered.
The same principle extends to every other category of AI-generated content. Factual assertions that bear on the outcome of a matter should be checked against primary sources, not assumed to be correct because the AI stated them confidently. Regulatory and statutory citations should be verified for current accuracy — AI training data has a cutoff date, and rules change. Procedural representations should be confirmed against the actual local rules of the court or the jurisdiction.
None of this is different in kind from what careful lawyers have always done when supervising junior work product. It is different in practice because AI output looks more finished than a first draft from a junior associate, and that polish creates a false impression of reliability that reduces the instinct to check. The supervision reflex that a rough first draft reliably triggers does not always fire when the output is fluent and well-formatted.
The Supervision Protocol Your Firm Probably Does Not Have
Most firms that have adopted AI use policies have addressed disclosure, confidentiality, and acceptable tools. Fewer have adopted AI supervision protocols — documented processes that specify what review is required for which categories of AI-generated content, who is responsible for that review, and what the standard of verification is before work product goes to a client or a court.
The absence of a protocol is itself a supervision failure. If different lawyers in the same firm are applying different standards to AI output — some verifying citations, some not; some checking factual assertions, some assuming accuracy — the firm does not have a supervision practice. It has individual habits of varying quality, which is a different and more dangerous thing.
A workable supervision protocol does not need to be elaborate. It needs to specify, for the categories of work the firm produces with AI assistance, what verification steps are required and who performs them. Citation verification. Factual accuracy checks. Currency of legal authority. Format and procedural compliance. Those categories cover the majority of the risk, and a one-page protocol that addresses them clearly is more useful than a lengthy policy that addresses disclosure while leaving verification to individual judgment.
The Discipline Behind the Supervision
Trial lawyers live by a discipline that applies directly here: never ask a question on cross-examination unless you already know the answer. The same instinct governs AI supervision — verify so thoroughly that no error slips through because you assumed the machine had it right. The machine does not know it is wrong. It has no instinct for the difference between what it knows and what it has generated. It cannot flag its own uncertainty in a way that reliably maps onto actual uncertainty. The only check on AI error is the lawyer who reviews the output — and that check works only if the review is designed to catch errors rather than to confirm that the output looks good.
Courts are not sympathetic to the argument that AI error is a mitigating factor. The lawyer signed the document. The lawyer filed it. The supervision obligation runs to the lawyer, and it runs whether the error was made by a junior associate, a contract attorney, or an AI system that produced a plausible-looking citation for a case that does not exist.
AI is not going anywhere; it will keep drafting, keep citing, keep sounding sure of itself. The only question is whether you supervise like the new boss or the old one. Get that right, and you won’t get fooled again.
If questions about AI supervision, quality control, or practice management are relevant to your firm or your clients, we are glad to have that conversation.
A Word About Silver Cain
Silver Cain PLC represents businesses in complex commercial and real estate litigation in Arizona and beyond. When Rebecca Cain and I founded the firm, we built it around direct partner involvement, senior trial-level judgment, and a cost structure that passes the efficiencies of serious AI use through to clients rather than burying them in the leverage model. Leon Silver and Rebecca Cain have spent decades handling high-stakes business disputes in Arizona and nationally. If the questions in this post are relevant to your business — or to the firms you retain — we are glad to have that conversation.
