By Jeffrey T. Donner, Esq.

September 11, 2026

For lawyers, the most troubling part of the recent decision in Douglas v. Deutsche Bank National Trust Co. is not that artificial intelligence hallucinated four nonexistent cases. Artificial intelligence can make mistakes. Lawyers know that. Courts know that. Technology companies know that. The troubling part is that, in September 2026, lawyers representing one of the world’s major financial institutions allowed those nonexistent cases to appear in an appellate brief without anyone verifying that the cases actually existed.

That should not happen. More precisely, at this point in the development of generative artificial intelligence, it cannot reasonably be treated as an understandable mistake attributable to unfamiliar technology. The danger has been known for years. The first nationally publicized episode involving fabricated AI-generated authority, Mata v. Avianca, Inc., occurred in 2023. Since then, courts, bar associations, continuing legal education programs, legal publications, and law firms themselves have repeatedly warned lawyers that generative AI can invent authorities, quotations, procedural histories, and propositions of law. By September 2026, ignorance is no longer a plausible explanation.

What makes the episode particularly disturbing is that the underlying professional obligation has almost nothing to do with AI. Lawyers have always had an obligation to verify their authorities. Long before ChatGPT, Gemini, Westlaw AI, Lexis+ AI, or any other generative system existed, a lawyer citing a judicial opinion was expected to read the opinion, understand what the court held, determine whether the case remained good law, and represent its holding accurately. That is not advanced appellate technique. It is among the most elementary disciplines of legal research and writing.

When I attended the University of Florida College of Law, nobody had to warn us about artificial intelligence because it did not exist in any commercially meaningful form. But the underlying rule was drilled into us nevertheless: do not cite a case for a proposition the case does not support. Do not turn dicta into a holding. Do not omit language that materially changes the meaning of the passage you are quoting. Do not characterize an argument made by a party as though it were the conclusion of the court. And certainly do not cite authority you have not actually read.

That discipline has remained with me throughout my career. A lawyer should have a healthy degree of apprehension about getting a citation wrong—not fear in a pathological sense, but professional respect for the consequences of telling a court what the law is. Opposing counsel may check the case. The judge or law clerk may check the case. More importantly, the court is entitled to rely on counsel’s representation that the cited authority exists and says what counsel represents it to say.

That is why an entirely fabricated case is qualitatively different from an ordinary research error. Misreading a complicated opinion can be negligence. Missing a later case can be negligence. Even experienced lawyers can disagree about what a holding means. But citing a judicial decision that literally does not exist reflects the absence of the most rudimentary verification step imaginable: opening the case.

The facts of Douglas make the problem difficult to minimize. Deutsche Bank’s lawyer acknowledged that she had used Google’s generative artificial-intelligence search tool to assist in locating case authority and had failed to verify the existence or accuracy of four authorities before filing the brief. The cases were imaginary. The D.C. Court of Appeals struck the entire brief and referred the matter to disciplinary authorities.

The attorney also advised the court that this was the first appellate brief she had filed. That fact may help explain how the mistake occurred, but it does not excuse it. Indeed, an inexperienced appellate lawyer should be subject to greater supervision, not less. Appellate briefing is precisely the kind of work in which an inexperienced lawyer should expect a senior lawyer to review the substantive argument, important authorities, record citations, and procedural posture before the brief is filed.

And that brings the case beyond individual error and into the subject of law-firm management.

Two other attorneys were listed on Deutsche Bank’s brief. The firm later told the court that it prohibited employees from using AI to draft legal documents and required citations to be verified regardless of their source. But the firm apparently did not explain what substantive supervision or review had occurred before the brief was filed. The Court of Appeals found that omission significant. It observed that if any of the firm’s attorneys had reviewed the brief before filing, the fabricated authorities would have been discovered.

That principle is also older than AI. Rules governing competence and supervision have long required lawyers and law firms to ensure that legal work is performed with appropriate knowledge, thoroughness, preparation, and oversight. A law firm cannot solve the problem merely by adopting an AI policy and circulating it internally. The policy has to be reflected in actual practice.

The necessary system is not complicated.

Before a significant brief is filed, someone must be responsible for verifying the authorities. That does not necessarily mean that every partner whose name appears in the signature block must independently pull every case. A law firm may divide responsibility among lawyers and trained staff and establish reasonable verification procedures. What it cannot do is have no reliable verification process at all.

In practice, a firm could adopt an embarrassingly simple rule: every cited judicial decision must be opened in Westlaw, Lexis, an official court database, Google Scholar, or another reliable source before filing; the lawyer responsible for the proposition must read enough of the decision to confirm that it actually supports the proposition; and somebody must be accountable for the final cite check. For particularly important dispositive motions or appellate briefs, there is nothing unreasonable about requiring the responsible lawyer to have the principal authorities collected electronically or printed and available during final review.

That was normal lawyering before artificial intelligence. It should remain normal lawyering now.

I use artificial intelligence in my own practice, extensively. Used correctly, it is an impressive professional tool. It can accelerate research, identify issues, organize facts, compare arguments, summarize lengthy materials, suggest lines of inquiry, critique drafts, and dramatically reduce the time required to get from a blank page to a serious piece of legal analysis. In a small practice, those efficiencies matter enormously.

But I do not treat AI as a judicial reporter.

When an AI system gives me a case, statute, rule, quotation, or factual proposition that matters, I verify it through the source material. If the case matters to the argument, I read the case. I check whether the quoted language actually appears in the decision. I determine whether the proposition is the court’s holding rather than a party’s argument or dicta. I look at later authority where necessary. That is not a special precaution required because AI exists. It is what a lawyer is supposed to do regardless of where the lawyer initially learned about the case.

My own experience has been that these systems have become materially more reliable as they have improved and as users have learned how to use them more carefully. But greater reliability does not change the lawyer’s duty. It simply makes the tool more useful. No matter how good the technology becomes, the responsibility for what appears in a filed brief remains with the lawyer.

The same caution applies to legal-specific AI products. Lawyers may understandably assume that an AI product embedded inside Westlaw or Lexis must be inherently safer because it sits on top of a trusted legal database. That assumption should not replace verification either. A fake case is obvious once somebody searches for it. A real case inaccurately described by AI can be considerably more subtle.

I have used legal-platform AI products, and I have not regarded the mere fact that they were attached to a premium legal database as a substitute for judgment. The relevant question is not whose logo appears on the screen. The relevant question is whether the lawyer has tested the tool, understands what it can and cannot do, and independently validates anything material before putting it before a tribunal.

That is also why I reject the developing narrative that AI necessarily favors only the largest law firms. In many respects, the opposite may be true. Properly used, AI can be an extraordinary equalizer. A small litigation firm does not have fifty associates available to summarize depositions, brainstorm issues, generate research leads, reorganize a record, or produce the first analytical pass on a complicated problem. Technology can provide leverage that once required an enormous institutional infrastructure.

That does not mean a small firm can magically reproduce every resource of a multinational law firm. It does mean that the gap has narrowed. An experienced trial lawyer equipped with strong technological tools, judgment developed over decades, and disciplined verification procedures can perform sophisticated work far more efficiently than would have been possible even ten years ago.

The essential word, however, is judgment.

Artificial intelligence can expand a lawyer’s capacity. It cannot assume the lawyer’s professional responsibility. A lawyer signs the filing. A lawyer represents to the tribunal that the authorities cited in it are real. A lawyer decides whether the case stands for the proposition asserted. A lawyer owes the duty of competence to the client and the duty of candor to the court.

There is, of course, a legitimate question about what discipline should follow when those obligations are violated. A lawyer who knowingly fabricates authority commits misconduct of the most serious kind. Where the conduct is negligent rather than intentional, proportionality matters, and professional discipline should take into account the lawyer’s state of mind, prior history, degree of supervision, remedial efforts, and the circumstances of the violation.

But the profession should stop treating fabricated authorities as quirky “AI mistakes.” Repeated, reckless, or grossly negligent submission of nonexistent authority should trigger serious professional consequences and a genuine inquiry into whether the lawyer is practicing with the competence the profession requires.

By September 2026, the bar for avoiding this problem is not high. It is extraordinarily simple: before citing a case, make sure the case exists. Before representing what it holds, read it. Before filing the brief, verify the citation.

The remarkable thing about Douglas is therefore not that artificial intelligence hallucinated. Machines make mistakes. The remarkable thing is that licensed lawyers allowed the machine’s mistake to become the court’s problem.

That is not a technology failure.

It is a lawyering failure.

And by now, there is no excuse for it.