Cite Checking to Find Hallucinated Cases Deemed Insufficient – Part II – Verification and Candor Are Expected

A North Carolina federal court recently held that reliance on cite checking software to disclose hallucinations prior to filing a document in court is not the equivalent of having an attorney, or a paralegal working at an attorney’s direction, verify citations as was required by the Court’s Standing Order.

In May, I published a blog about  Williams v. Honl, 348 Or. App. 505 (Apr. 22, 2026), titled Cite Checking to Find Hallucinated Cases Deemed Insufficient.  The Oregon court wrote: “[S]o that the clear does not become cloudy, we state the obvious: Using generative artificial intelligence to generate legal briefs and then simply cite-checking them bears no resemblance to the competent practice of law….When lawyers trade reflection for automation, they surrender the very quality that makes their words worthy of belief. Although cite checking is, of course, an important part of producing reliable, competent briefs, it is not the type of work that requires a law degree. Law is a profession….” (cleaned up; citation and quotations omitted; emphasis added).

A similar result was reached in Smith v. Polk Co., 2026 WL 2280990 (W.D. N. Car. Aug. 7, 2026). After the court issued an order directing Plaintiff to show cause why she should not be sanctioned for use of artificial intelligence, counsel for Plaintiff responded by letter. Plaintiff apologized for inadvertently failing to file an A.I. certification required by the court. Additionally:

Counsel represented that she “does not use AI to conduct legal research,” and that “[t]o ensure the accuracy of citations … [she] runs briefs through LexisNexis brief analysis” and fixes any error that appears “alarming.” … Counsel further represented that she “has AI embedded in her practice” but that her “AI program cannot conduct research and cannot hallucinate.” … Regarding the fabricated quotations, counsel for the Plaintiff says that she “must be the origin,” that “she made the errors, accepts full responsibility, attempted to properly verify her work and fell short,” and that “AI cannot be blamed.” … Counsel also represents that she has now “made changes to her practice” and “has added additional time to more thoroughly vet her work.”

Id. at *10 (emphasis added). The court was not satisfied:

The Court is deeply troubled by the errors previously identified in the Plaintiff’s filings, and counsel’s letter does little to allay the Court’s concerns. Counsel indicates that she relies on a LexisNexis program for citation verification rather than having an attorney, or a paralegal working at an attorney’s direction, verify citations as required by the Court’s Standing Order. Moreover, counsel has not explained the origin of the fabricated quotations other than to accept responsibility for them and assert that they were not produced by an artificial intelligence program. How such fabricated quotations could have appeared in the Plaintiff’s filings at all therefore remains a mystery. Finally, while counsel asserts that the artificial intelligence program embedded in her practice cannot hallucinate, she has provided no corroboration for that assertion, nor has she even provided the name of the artificial intelligence program that she uses. As a result, the Court finds that counsel’s response has fallen well short of the Court’s expectations. Nevertheless, because counsel has accepted responsibility for the errors in the Plaintiff’s filings, the Court will discharge the Show Cause Order. Counsel for the Plaintiff is cautioned, however, that any similar shortcomings in future filings will result in sanctions.

Id. (emphasis added).

The court in Barteca Holdings LLC v. Tacobarn Newtown LLC, __ F. Supp. 3d __, 2026 WL 2253785 (D. Conn. Aug. 4, 2026), issued a show cause order concerning use of GenAI.  It cataloged erroneous citations, quotes that did not exist, and mischaracterizations of law. After hearing the attorney’s explanation, it wrote:

On June 25, 2026, the Court held a show-cause hearing, at which Attorney Miller reiterated those representations and further explained the circumstances under which the AI-generated material was incorporated into his briefing. Attorney Miller clarified that this matter is the first and only matter in which he utilized artificial intelligence. As for his writing process, Attorney Miller explained that he first wrote the brief himself conventionally. Then, in an attempt to improve it, he used a tool called Open Law, which he paid for, for suggestions on arguments and additional cases. He also used Claude and Chat GPT for additional suggestions. Concerned about the risk of AI hallucinations, Attorney Miller then cross-checked the citations in the brief to determine whether they existed and were cited correctly, including by Shepardizing every case in the brief through Lexis. However, he did not adequately verify that each “pin cite” referred to the correct page or that the quotations included in his brief were accurate. [emphasis added].

The court wrote that “merely incorrect legal statements” do not violate Fed.R.Civ.P. 11. However, once a fabricated citation is identified, the court considers two factors:

1) verification and inquiry (“Whether counsel [or the unrepresented party] conducted a reasonable, human-based verification of every cited authority before filing. A reasonable inquiry requires more than reliance on an automated tool; it demands independent confirmation through recognized primary legal sources. The signature of an attorney [or the unrepresented party] certifies human diligence, not mechanical output.”)(emphasis added; citation omitted); and,

2) candor and correction (“Whether counsel [or the unrepresented party] promptly disclosed the use of AI and corrected the record once inaccuracies were discovered. Candor after filing weighs heavily in mitigation; concealment or minimization aggravates the violation. The duty of candor extends beyond the moment of signature, it continues for as long as the filing remains before the Court.”)(citation omitted).

The Barteca court wrote: “Applying the framework discussed above, the Court finds that the first factor—verification and inquiry—weighs in favor of sanctions because Attorney Miller failed to conduct a reasonable human verification of the authorities cited in the challenged filings before submitting them.”  It stated that “Rule 11 imposes an obligation on attorneys to, at a minimum, read the cases they cite to the Court.” [citation omitted].

While the second factor—candor to the court—weighed in the attorney’s favor, “regret and apologies are not necessarily enough to avoid the imposition of sanctions for the submission of non-existent legal authority.” [citation omitted].

After consideration of all factors, the court imposed a sanction of $3,500 “together with a referral of this matter to the Grievance Committee….”

For general guidance, see Annot., 110 A.L.R.7th Art. 3 (Originally published in 2026), titled “Sanctions and Remedies Where Parties Submit Citations Hallucinated by Artificial Intelligence Programs—State Cases.”

This blog was initially posted on  Electronic Discovery Reference Model.