A judge’s order rife with AI-induced errors gives state and federal courts serious food for thought as they finalize guidelines governing whether and to what extent judges and their law clerks can use generative artificial intelligence tools in drafting court orders and opinions.
Last summer, U.S. District Court Judge Henry T. Wingate in the Southern District of Mississippi issued a temporary restraining order against the state in Jackson Federation of Teachers v. Fitch. The TRO entered in the civil rights case referenced allegations and parties not in the lawsuit, nonexistent declarations by parties and witnesses, and language not found in the state law being challenged by the plaintiffs.
When made aware of those facts in a motion to clarify by the state, Wingate completely removed the original TRO from the docket and replaced it with a corrected order.
But the matter did not end there. Word had gotten out that the errors in the original TRO were produced by a generative AI tool that one of Wingate’s law clerks used in drafting the document. That news prompted calls from a high-ranking U.S. senator for an explanation, and a panel of the 5th U.S. Circuit Court of Appeals last month heard arguments as to whether Wingate should be removed from the case.
As state judiciaries and the federal courts finalize guidelines for the use of generative AI tools by judges and other court personnel, attorneys like Boston lawyer Graydon H. Sommer urged caution.
“Courts should base policy on demonstrated reliability,” said Sommer, whose practice includes a focus on AI law. “Improved performance may justify different uses, but it should not change who is responsible for the accuracy and reasoning of a signed order.”
Class A screwup
The plaintiffs in the Fitch case had sued the state of Mississippi in June 2025 to block the enforcement of H.B. 1193, an anti-DEI bill that had just gone into effect.
On July 20, 2025, Wingate granted the plaintiffs’ request for a TRO.
Two days later, the state filed a motion to clarify or correct the order, alleging the TRO “identifies incorrect plaintiffs and defendants; recites allegations that do not appear in the operative complaint and/or are not supported by record evidence; identifies, as quoted excerpts, certain terms that do not appear in the language of H.B. 1193; and relies upon the purported declaration testimony of four individuals whose declarations do not appear in the record for this case.”
State attorneys alleged that, in response to their motion for clarification, Wingate removed the original TRO from the docket and replaced it with a corrected TRO that was signed on July 22, 2025, but backdated to July 20.
Later, the state appealed Wingate’s issuance of a preliminary injunction sought by the plaintiffs to the 5th Circuit.
Prior to oral argument, a 5th Circuit clerk sent a letter advising counsel to be prepared to address “whether, especially in light of the use of AI by the district court, this matter should be sua sponte assigned to a different district judge.”
Sen. Charles E. Grassley, R-Iowa, who chairs the Senate Committee on the Judiciary, caught wind of what had transpired and sent Wingate a letter last October expressing concern that the court had used artificial intelligence to prepare the original TRO.
“These do not appear to be simple slips of the pen or mechanical oversights, but substantive errors that undermine confidence in the Court’s deliberative process,” Grassley wrote.
Wingate responded with a letter of his own in which he explained that a law clerk had utilized an AI tool “strictly as a foundational drafting assistant to synthesize publicly available information on the docket.” He wrote that the docketed order “was an early draft that had not gone through the standard review process.”
What occurred in the Wingate matter mirrors the issues that lawyers encounter incorporating AI technologies in their own practices, said McLane Middleton attorney John F. Weaver.
“There are reasons why firms have developed training programs and AI use policies for their attorneys and staff,” he noted.
Consequences, consequences
During oral argument in the Fitch case last month, the state’s attorney, Anthony M. Shults, argued for reassigning the case to a different judge.
“We believe the court’s actions would lead any reasonable observer to question whether it has the ability to afford this case due care to accuracy going forward,” Shults told the 5th Circuit panel.
Shults also argued that Wingate’s amended TRO included at least one AI hallucinated case citation from the original order.
Circuit Judge Jerry E. Smith asked the plaintiffs’ attorney, Robert B. Duff, whether he thought it was a “serious matter” if after all that had occurred with the original TRO there would still be an “AI hallucination” in Wingate’s amended order.
Duff responded: “It is serious, but it’s not as if AI told Judge Wingate to grant the TRO.”
But in the 5th Circuit’s recording of the oral argument, Smith sounded exasperated by Duff’s attempt to downplay the seriousness of Wingate’s issuance of an order plagued by AI hallucinations.
“Are you saying that we can be confident that that didn’t infect his ultimate reasoning in the case? How can we be confident about that?” Smith asked.
Smith honed in on the proper remedy, asking counsel if the panel were to decide to assign a new judge, whether the better course would be to vacate Wingate’s preliminary injunction and let the new judge address the appropriateness of injunctive relief on a clean slate.
Fitch crystallizes for courts considering the proper use of AI tools the issue of what remedies might be in order when a judge issues a ruling tainted by faulty output.
According to Sommer, the relief available to a party from a court order that contains AI-induced errors would depend on a variety of factors, including the jurisdiction, the procedural posture of the case, the nature of the error, and applicable deadlines.
Weaver said it “might make sense that, as a matter of policy, the litigants get another chance to argue their case, because this type of malfeasance does introduce mistrust into the system.”
Meanwhile, Davis Malm civil litigator Anthony T. Panebianco said he believes counsel have an obligation to notify the court when faced with an order that appears to include hallucinated AI output, even when the ruling may be favorable to their side.
“I would reach out to opposing counsel and say, ‘I think this is wrong; there’s a mistake here,’ and ask whether we want to reach out together to bring it to the court’s attention,” Panebianco said.
