Two years ago, few lawyers would have thought to ask this question: Should you cite-check a court decision in your favor for AI hallucinations? A case now pending in the U.S. Court of Appeals for the Fifth Circuit has raised that question and opened up an opportunity for lawyers to start thinking hard about how to handle this new issue.
The case is Jackson Federation of Teachers v. Fitch, a challenge to a Mississippi law. In July 2025, the U.S. District Court for the Southern District of Mississippi entered a temporary restraining order blocking parts of the statute. The order was, to put it charitably, not in good shape. It referred to parties who had never been in the case. It described allegations no one had made. It quoted declarations from witnesses who do not exist. The state raised the problems with the court, and the order came off the public docket and was replaced with a corrected version. The corrected version still included a hallucinated citation. The court described the difficulties as clerical and declined to say much more. The fuller explanation came later: A law clerk had used a generative AI tool while working on an early draft, and that draft had gone out without the layers of review the chambers ordinarily applies. How exactly it escaped detection has never really been explained.
That might have been the end of it. It was not. The preliminary injunction that followed did not disclaim the TRO but rather incorporated the TRO’s reasoning. At oral argument, the Fifth Circuit asked both sides to be prepared to address whether the case should be reassigned to a different district judge. At argument, one panel member pressed counsel about the seriousness of a hallucinated citation surviving even after the problems with the original order had come to light.
Counsel defending the injunction had a good answer: AI did not decide the case. The judge held an evidentiary hearing, heard the evidence, and made the decision himself. Whatever was wrong with the TRO’s prose, the ruling was the judge’s ruling. Errors in the document did not necessarily mean errors in the decision. That argument may well carry the day, but it seemed more likely the injunction would be reversed.
The case raises a hard question. Would the plaintiffs be in a better position today if they had been the ones pressing hardest for a clean order? Suppose they had joined the state's motion rather than resisting it. Suppose they had asked the court to reissue the order in full, said plainly that they wanted every citation verified before anything went up on appeal, and then checked the corrected version themselves, line by line, the way they would check an order that had gone against them. The record before the Fifth Circuit would look very different now. There would be no live argument that the injunction they won rests on a document nobody has fully accounted for.
But the risks and costs of exposing a court’s mistakes are obvious. Nobody wants to tell the judge who has just ruled for them that the opinion has problems. You have persuaded a court, and your instinct—a sound instinct, most of the time—is to take the win, say thank you, and move on. Raising your hand to say the order needs work can look like ingratitude or like a lack of confidence in your own victory, and there is a real risk of turning a favorable order into a less favorable one on a second pass. That calculus gets harder still as the errors get subtler. A quotation from a phantom witness is easy to identify and correct. A quotation that is slightly off, or a case cited for a proposition it does not quite hold, is the sort of thing lawyers would be tempted to let go.
There is no clean rule here. But two things seem to be true at the same time. The first is that the old division of labor—the losing side polices the order, the winning side enjoys it—assumed that the errors in an opinion would be the ordinary humankind. Confident, fluent, entirely fabricated citations are a different animal, and they are dangerous to the party who benefits from them in a way that a garden-variety mistake never was. The second is that our appellate courts are only beginning to work out what to do about all this. Until they do, the safest assumption is that a hallucination in your favorable order is a defect in your judgment, and one you will be asked about later.
This column has previously suggested that AI is a capable assistant and a poor substitute for professional judgment. The lesson has mostly been aimed at what we file. Jackson Federation suggests the same lesson now runs in the other direction, toward what a lawyer receives from the court. Sometimes, the gift horse might need to be looked in the mouth.