THE AUTHOR:
Prof. Barry Appleton, New York Law School
On October 22, 2026, the United States Court of Appeals for the Fifth Circuit vacated a preliminary injunction and directed that the case be reassigned to a different district judge. The restraining order underlying it had named eleven individuals and organizations that had never been parties, named two officials who were not actually sued, referenced declarations from four individuals who do not appear in the record and may not exist, and attributed to a Mississippi statute terms that appear nowhere in it. The judge later acknowledged that artificial intelligence (“AI”) had been used in preparing that order.
The injunction was vacated for legal error, and the court said so plainly. What turned a vacatur into reassignment, a power the majority called extraordinary and rarely invoked, was largely the thirteen weeks that followed the first order. That sequence is the part of Jackson Federation of Teachers v. Fitch, No. 25-60496 (5th Cir. Oct. 22, 2026) that transfers to a tribunal, because every temptation the judge gave in to presents itself to an arbitrator who finds a fabricated citation in a draft award at two in the morning. The reasoning is Judge Smith’s, for a majority; Judge Haynes concurred in the disposition only.
The Thirteen Weeks
Judge Henry Wingate issued the temporary restraining order (“TRO”) on July 20, 2025. Two days later, the defendants moved to clarify, listing the inaccuracies. That same day, without explanation, the district court removed the original order from the docket and replaced it with what the majority called a “seemingly corrected” version, backdated to July 20.
The defendants then asked for something modest: restore the original to the docket as a superseded order so the record would be whole, and have the court explain the inaccuracies. The court refused, characterizing the fabricated parties, hearings, and facts as “clerical.”
The opinion records that after accusations of AI use made national headlines, the judge acknowledged the errors, attributing them to a law clerk’s use of AI in drafting an earlier version of the order, docketed by mistake. It also records that he admitted to using AI to prepare the order himself. The sequence deserves a note: the opinion says the admission came fifteen days after a United States Senator accused the court of using AI, while its own footnotes date the judge’s letter four days before the Senator’s. The released correspondence supports the text rather than the footnote. Senator Chuck Grassley wrote on October 6, 2025; Judge Wingate’s reply is dated October 21, released two days later.
An opposing party saved the record. In a footnote, the majority observed that it was “grateful—and, indeed, lucky—that the defendants provided a copy of the original TRO as an exhibit in their July 28 motion . . . given that Judge Wingate kept it from public view.”
“Clerical” Is a Claim, Not a Description
The judge’s letter explained that the errors were clerical because they did not relate to “core legal reasoning or the ultimate judgment on the merits,” but instead to a “failure to confirm the accuracy of certain information.”
The majority was unpersuaded: “It is not clear to this court how so many mistaken facts could not at least marginally ‘affect’ ‘core legal reasoning’ or support for the judgment on the merits.” It found that the judge had “dug in his heels,” offering no explanation as to why apparently irrelevant material had been included in the first place, or why the order was justified once that material was gone.
This is the part that travels. Calling an error clerical is an assertion about the relationship between the offending material and the reasoning itself. The distinction it invokes is real enough, since a failure to verify is not a failure to reason. But it states a conclusion about the document rather than describing it; it does not bind the reviewer, who will test it. A decision-maker who makes that claim without disclosing how the material got there is asking to be believed on precisely the point he has declined to explain.
How the Fake Citation Was Caught, and How It Was Missed
Opposing counsel read the order against the record within two days and caught the first batch. One escaped. The “seemingly corrected” order still cited Cousins v. School Bd. of City of Norfolk, 503 F.2d 422, 426-27 (4th Cir. 1974), a case that does not exist. It survived the judge’s own revision and survived incorporation by reference into the preliminary injunction, and was discovered only at oral argument in the Fifth Circuit.
The method that caught it is the practical takeaway. The majority traced the reporter reference: 503 F.2d 422 falls inside United States v. Anderson, 503 F.2d 420 (6th Cir. 1974), and the pinpoint pages land in Avis Rent A Car Sys., Inc. v. United States, 503 F.2d 423 (2d Cir. 1974). In a review of an arbitration award in Association des ressources intermédiaires d’hébergement du Québec (ARIHQ) c. Santé Québec, 2026 QCCS 1360, the Quebec Superior Court did the same thing, finding that the numeric reference behind each hallucinated judgment led to a completely different decision, and that the one real decision in the set had been cited for a proposition it does not support. Neither case required expert evidence. Someone pulled the citations.
The uncomfortable part is that a correction cycle, an opposing party, a district judge, and a round of appellate briefing all missed the one that got through, in a case where everyone was already looking.
What Arbitration Does Not Have
An arbitrator in this position faces the same temptations with fewer restraints, though not the restraints usually assumed.
There is no public file. Most commercial awards are never published, so an arbitrator’s reliance on a fabricated authority ordinarily never becomes public. ARIHQ is the first decision anywhere to have annulled an award for the arbitrator’s own AI use, and so far as I know still the only one. What arbitration does have is an institutional file, and it is less forgiving than a court docket: under the 2026 ICC Rules, the Secretariat notifies the award and keeps an original, the parties may obtain certified copies, and a tribunal cannot quietly substitute a notified award. In an administered arbitration, the quiet fix is largely unavailable.
There is no appellate supervision. A court of appeals can vacate and reassign under its supervisory power, in public, on the record. An institution can act: under Article 16(2), the ICC Court may initiate replacement proceedings on its own initiative where an arbitrator may not be fulfilling the arbitrator’s functions in accordance with the Rules. But outside an administered arbitration, removal runs through a challenge or an annulment, initiated by the parties and paid for by the parties.
There is no merits review. In Quebec, a court hearing an annulment application may not examine the substance of the dispute. The respondent in ARIHQ was right that AI use appears nowhere among the enumerated grounds. The court said so, and annulled anyway, because the arbitral procedure followed was not the procedure that applied. The hallucinated authorities were the award’s only legal support, and the court concluded that the arbitrator had delegated his authority and abdicated his role in reviewing the result. The award was annulled with costs, and the parties were ordered to choose a new arbitrator within sixty days.
The Correction Power Will Not Reach It
Here is where the instinct to fix an award quietly meets the architecture.
Article 39(1) of the ICC Rules permits a tribunal, on its own initiative and after seeking the parties’ comments, to correct “a clerical, computational or typographical error, or any errors of similar nature,” provided the draft correction reaches the Secretariat within 45 days of notification. Article 39(2) gives a party 30 days to apply. Article 33 of the UNCITRAL Model Law turns on the same categories.
The case for reaching for it is better than it looks. A citation inserted by a drafting tool, never read by the tribunal and playing no part in the chain of reasoning, is on one view indistinguishable from a mistyped reporter number: an error in the expression of the award rather than in the tribunal’s thinking. The orthodox question is whether the correction would alter the decision.
The difficulty is evidentiary, and in practice it is fatal. A tribunal cannot establish that a fabricated authority was decorative without explaining how it got there, which means disclosing the tool and the stage at which it was used. Reaching for the slip rule therefore concedes the very thing a quiet correction was meant to avoid. Where the authority was load-bearing, the characterization fails on its own terms. That is the line ARIHQ draws: the court reserved the case of minimal AI use, or use touching a less important question, where the weighing might produce a different result, and annulled because these citations sat at the heart of the reasoning.
Two further points cut against the quiet fix rather than for it. First, a tribunal’s mandate does not end as cleanly as functus officio suggests. Under Article 32(3) of the Model Law, it terminates subject to Articles 33 and 34(4). Article 34(4) lets a court seized of a setting-aside application suspend those proceedings, on a party’s request, so the tribunal can resume and eliminate the grounds. ICC Article 39 runs to the remission of awards, and Article 39(5) governs what follows once a court remits one. The route back runs through the parties and the court, not a reissued document. Second, a formal correction request extends the challenge clock rather than consuming it: under Model Law Article 34(3), the three-month period runs from receipt of the award, or, where an Article 33 request was made, from its disposal. It is the informal substitution, made without any request, that burns a party’s time, and that is no correction at all.
The Morning After
Since the Silicon Valley Arbitration and Mediation Center issued its AI guidelines in 2024, the arbitration community has argued over whether arbitrators should disclose their use of these tools before writing an award. I have taken a position in that debate on what belongs inside the ICC’s AI architecture. Fitch is about a different question, and the instruments are thinner on it: what a decision-maker owes the parties once something has already gone wrong.
The protocol I would draw from it is short. Tell the parties, in writing, all of them at once. Do not substitute one document for another without explanation. Describe the problem rather than characterizing it, and let the parties and any supervising authority supply the characterization. Identify the affected passages, state what was used and at what stage, state what has been verified since and how, and invite submissions on what should follow, including whether the tribunal should continue.
An arbitrator who discloses a hallucinated citation promptly has a procedural problem, and it may be a serious one. An arbitrator who reissues the award quietly and calls it clerical has that problem, plus a credibility problem, plus two inconsistent documents in circulation and no explanation for the difference. The first may keep the award. The second is arguing against the ARIHQ outcome with the Fitch judge’s record.
According to a note on the ICC Commission’s October 3, 2026 meeting, the ICC’s Task Force on Artificial Intelligence in Dispute Resolution has circulated a draft protocol on arbitrators’ use of AI, sorting permitted assistance into tiers by the nature of the work. Sorting by function is the right starting point, and it begins to answer the question the field has been asking. Fitch raises the one the instruments have not reached: what an arbitrator does the morning after. That is the gap the next instrument should close.
