Summary:
Generative artificial intelligence has now reached the Eastern District of North Carolina in a published sanctions decision, and the resulting opinion should be required reading for every practicing attorney.
Although Fivehouse v. United States Department of Defense is an Administrative Procedure Act case involving veterans' health care—not bankruptcy—the lessons apply equally to bankruptcy practice, civil litigation, and virtually every other area of law. More importantly, because the transcript of the show cause hearing remained under embargo for several weeks before becoming publicly available, the court's final amended order is best understood when read alongside that transcript. Together, they tell a fuller story than either document alone.
The litigation began ordinarily enough. Plaintiff challenged decisions affecting TRICARE coverage under the Administrative Procedure Act. During the litigation, the Assistant United States Attorney filed several briefs opposing the plaintiff's motions. The plaintiff eventually pointed out that one of those briefs contained fabricated quotations, inaccurate parenthetical descriptions, and misstatements of what several Fourth Circuit decisions actually held. The court independently reviewed the cited authorities—including Ohio Valley Environmental Coalition v. Aracoma Coal Co., Dow AgroSciences, LLC v. National Marine Fisheries Service, and Sierra Club v. United States Department of the Interior—and concluded that many of the quoted passages simply did not exist.
Initially, however, that was not the explanation given.
When first ordered to respond, counsel stated that the errors resulted from "the inadvertent filing of an unfinalized draft document" and asked either to strike the erroneous quotations or substitute what he described as the intended version of the brief. Notably absent from that explanation was any mention of artificial intelligence.
The Hearing Told the Rest of the Story
The show cause hearing dramatically changed the picture.
Under oath, after the court repeatedly asked for a complete explanation, the AUSA testified that he had accidentally overwritten an earlier draft of the response, panicked upon discovering the mistake shortly before the filing deadline, and used generative artificial intelligence to recreate the brief. He further acknowledged that he failed to verify the AI-generated quotations and citations before filing the document. When the court walked through the fabricated quotations one by one, he agreed that the problems in that brief all stemmed from the same failure to verify the AI's output.
The hearing transcript also explains why the final order places such emphasis on candor. The court noted that the attorney's earlier written response omitted any reference to AI, despite later acknowledging under oath that AI had generated the offending material. The judge repeatedly afforded him opportunities to explain how the fabricated authorities came to appear in the filings before the complete explanation finally emerged.
The Court's Holding
The court found by clear and convincing evidence that the AUSA knowingly submitted a brief containing fabricated quotations and misstatements of case holdings after failing to verify AI-generated content. It concluded that this constituted an abuse of the judicial process warranting sanctions under the court's inherent authority. Rather than relying on Rule 11, the court held that its inherent authority was the better vehicle because the legal arguments themselves were not frivolous—the problem was the deception practiced on the court.
Perhaps the most important sentence in the opinion is also its simplest:
"The court should hold the attorney accountable because he is responsible for the authorities he presents to the court, regardless of how they came to appear in the brief."
That principle deserves to be quoted often.
The problem was not ChatGPT, Claude, Gemini, Copilot, Westlaw AI, Lexis AI, or whatever generative tool may exist next year.
The problem was filing legal authorities without independently verifying them.
Credit Where It Is Due
One aspect of the opinion deserves recognition.
The leadership of the United States Attorney's Office for the Eastern District of North Carolina responded with a level of candor and institutional accountability that reflects well on the office. According to both the hearing transcript and the final order, United States Attorney Ellis Boyle immediately involved the Department of Justice's Office of Professional Responsibility and General Counsel, issued an office-wide memorandum emphasizing the duty of candor to the court, directed attorneys always to verify AI-generated quotations against the original authorities, and required additional professional responsibility training for every attorney in the office. He also personally apologized to the court during the hearing.
That level of accountability has not always characterized governmental responses when AI hallucinations have surfaced in litigation. In some other cases involving government lawyers, the initial explanation has focused on inadvertent error or omitted any acknowledgment that AI played a role. Here, by contrast, once the matter reached the show cause stage, the leadership of the Eastern District of North Carolina acknowledged the seriousness of what had occurred, accepted institutional responsibility for preventing a recurrence, and implemented corrective measures. Those actions almost certainly influenced the court's decision to impose only a public reprimand rather than the more severe sanctions it expressly stated would otherwise have been justified.
Commentary
The sanction itself is noteworthy. The court observed that the misconduct, the lack of candor, and the attorney's position of public trust would ordinarily warrant significant sanctions. Nevertheless, because the attorney had already suffered substantial professional consequences—including his separation from the Department of Justice after the hearing—the court concluded that a public reprimand was sufficient to vindicate the integrity of the judicial process.
For me, however, the broader lesson is a personal one.
Readers of this blog know that I increasingly use AI tools to help organize research, summarize opinions, and identify issues for further analysis. They are extraordinarily useful. They also make mistakes that are often remarkably persuasive because they sound exactly like something a court might have written.
This case is an important reminder that artificial intelligence is a research assistant—not an editor, not a law clerk, and certainly not a substitute for reading the cases ourselves.
The duty to verify has always existed. We have always been responsible for checking work produced by associates, summer clerks, paralegals, treatises, and computerized research services. Generative AI simply joins that list of tools whose work must be independently confirmed before it reaches the court.
That reminder applies to me every bit as much as anyone else. Whenever I use AI to assist in preparing a post for this blog, the underlying opinions and authorities still have to be read and verified. If an AI tool misquotes a case, invents a quotation, or subtly overstates a holding, the responsibility does not belong to the software. It belongs to the lawyer who publishes it or signs the brief.
Technology is changing rapidly.
Our professional obligations are not.
Blog comments