Lawyers’ Bar Journal Article Discussing Their AI-Hallucination Errors Doesn’t Entirely Satisfy Judge, but …

From Doe v. Univ. of N.C. Sys., decided today by Chief Judge Martin Reidlinger (W.D.N.C.):

On November 10, 2025, the Court ordered the Plaintiff’s counsel to show cause as to why they should not be sanctioned for the many errors in the documents they submitted to the Court, including: (1) citing two cases that do not appear to exist (i.e., hallucinated citations); (2) quoting material that does not exist in the cases purportedly quoted (i.e., hallucinated quotations); (3) mischaracterizing holdings of cited cases; (4) citing cases that have no bearing on the proposition for which they were cited; and (5) failing to provide pinpoint citations and information identifying the courts from which opinions issued.

On November 19, 2025, during a hearing on the show cause order, the Plaintiff’s counsel admitted these errors and explained that, in large part, the errors resulted from misuse of artificial intelligence software (including a misunderstanding of how to properly use artificial intelligence) and a failure to verify the outputs of artificial intelligence software, even though the Plaintiff’s counsel had signed certifications stating that they had verified every statement and every citation in the documents they submitted to the Court. The Plaintiff’s counsel expressed remorse and offered to write an article for the state bar journal explaining their errors and the potential pitfalls of misusing artificial intelligence. The Court agreed that such an article—”an article that essentially says, we really screwed up and we almost got put under the jail, don’t fall into the pit that we did”—could help other lawyers wake up to the seriousness of attorney misuse of artificial intelligence. The Court has refrained from discharging the show cause order pending counsel’s preparation of the article.

The Plaintiff’s counsel submitted their proposed article to the North Carolina State Bar Journal, and it has now been published. Fred W. DeVore III and Rob Wilder, Guarding Against AI Errors: Ethical Risks for NC Attorneys, N.C. State Bar J. 1, 8-12 (Summer 2026) (hereinafter “the Article”). Now before the Court is the issue of whether this Article is sufficient to purge the show cause order and the proposed contempt/sanctions that arose from counsel’s errors.

The Court is compelled to express its disappointment in that the Article falls short of the Court’s expectations based on what counsel proposed at the November 19, 2025 hearing. The Plaintiff’s counsel were less than fully forthcoming in the Article about the scope of their errors in this matter. The Plaintiff’s counsel had submitted five documents—two motion briefs and three briefs in opposition—with citation errors of all sorts. After the Defendant’s counsel brought attention to some of the errors, the Plaintiff’s counsel used their two reply briefs to address their errors rather than further their client’s case. As a result, the Plaintiff’s counsel’s errors undermined their own advocacy, hindered the work of their opposing counsel, and wasted significant time and resources of the Court. Yet, the Article fails to fully acknowledge the impact of such errors, both on the litigation itself as well as their reputation as attorneys. If the purpose of the Article is to keep other attorneys from “falling into the pit,” one must accurately describe “the pit.” Counsel’s efforts failed to fully achieve that objective.

In addition, counsel’s Article reflects a failure to fully appreciate the severity of counsel’s errors. The Plaintiff’s counsel recount their experience in a mere three paragraphs, accounting for a small fraction of the article. In doing so, their errors were unduly minimized. For example, one attorney writes: “The cases I cited were correct, but the quotes were hallucinated. I checked the cases but not the quotes.” A cited case is not “correct,” nor has a case been “checked,” unless the attorney has verified that the case contains the quote or supports the proposition for which the case is cited. When an attorney presents a document to the Court, an attorney certifies “to the best of the [attorney’s] knowledge, information, and belief, formed after an inquiry reasonable under the circumstances” that the document complies with the substantive provisions of Rule 11 of the Federal Rules of Civil Procedure. An inquiry that does not involve verifying quotes is never reasonable under the circumstances.

In another instance, one of the attorneys writes that he “signed a certification that [he] did not use AI to write the brief, which was true, but [he] failed to verify the authenticity” of sources that he received from other attorneys, i.e., attorneys who were not attorneys of record in the case. That is not, however, what he certified. This Court’s Standing Order on the Use of Artificial Intelligence, requires every statement and every citation to an authority contained in a document to be “checked by any attorney and/or a paralegal working at his/her direction … as to the accuracy of the proposition for which it is offered, and the citation to authority provided.” The Standing Order further provides that no artificial intelligence may be used in the research for the preparation of a document except for the “artificial intelligence embedded in the standard online legal research sources Westlaw, Lexis, FastCase, and Bloomberg.” The attorney signed a certification that he complied with those requirements, which was not true. He submitted a brief that contained citations to artificial intelligence hallucinations—i.e., to cases that do not exist.

For these reasons, the attorneys’ minimization of their errors falls short of the Court’s expectations. The Court recognizes that there are length limitations in the preparation of such an article, and that a portion of that length is taken by providing context for the central points. However, the purpose of the Article was for such exemplary attorneys to explain how artificial intelligence can be a danger to those practitioners who do not fully understand how to use it. By minimizing their misconduct, they missed the opportunity to demonstrate to the bar how easily one can commit serious transgressions if one is insufficiently familiar with the operations of the various artificial intelligence tools that are now available. {The Court also notes that the Article unironically concludes with a purported quotation that is neither cited nor attributed to any speaker: “Someone once said, ‘Make AI a tool for humans, not a reason to lose what makes us human.'”}

Having fully vented the Court’s disappointment, it is noted that offending counsel have a long history of exemplary conduct before this Court. Their errors primarily reflect the fact that they attempted to use the newest tools of the trade without actually understanding those tools or the dangers they present. For these reasons, notwithstanding the disappointing shortcomings of counsel’s Article, the Court finds their efforts and their expressions of repentance are made in good faith and as such are sufficient to discharge the Order to Show Cause.

You can read the lawyers’ article for yourself here. I will say, to their credit, that they did publicly and in some detail admit to pretty serious errors, in an article that will be read by hundreds of their colleagues; that may be part of the “expressions of repentance” that helped the court decide against imposing sanctions. That does show significant contrition, I think, even if the judge is correct that the admissions weren’t fully accurate. Here are the key paragraphs:

When attorneys fail to verify AI-generated output, they risk submitting fiction to the court. The consequences can be severe. Here is what happened to us.

Fred’s Story: I represented a client in federal court. I prepared a summary judgment brief and consulted with two well-respected attorneys who were experts regarding the legal issues involved. They volunteered to review my brief and make suggestions on its content. The attorneys then collectively sent me several paragraphs of compelling arguments, complete with persuasive quotations and citations. I gladly incorporated them into my brief.

There is no worse feeling than receiving a copy of the defendant’s brief, where the defendant informed the court that the quotations and citations I had incorporated were hallucinations. I had signed a certification that I did not use AI to write the brief, which was true, but I failed to verify the authenticity of the supplemental authority I was given. This violated my certification. The fault does not lie with the consulting attorneys because my duty to verify the authorities is non-delegable. My penance was to resubmit a corrected brief and to warn the Bar of the consequences of non-approved AI sources. The moral is unavoidable: delegation does not eliminate responsibility.

Rob’s Story: I am co-counsel with Fred on the same case in federal court. We agreed to take over the case with very little time left for discovery. We had to review a large volume of discovery documents and become deeply familiar with the case and regulatory law in what was, for me, a new area of law. I felt I was behind on technology and hired a consultant to help me get up to date with AI. While writing briefs on dispositive motions, I relied on a commercial legal research program’s rendition of several quotes supporting the points I made in the brief. The cases I cited were correct, but the quotes were hallucinated. I checked the cases but not the quotes. Again, I was the person who certified that I had authenticated the cited information. Before this incident, I was very confident I was using legal AI as effectively as possible, only to be shown I was wrong. Fortunately, the consequences of my mistake have been rectified. Hopefully, my experience can help others avoid the same embarrassment, anxiety, and frustration.