A federal judge used AI to prepare a court order, and the case has now been taken away from him. On 22 October 2026 the US Court of Appeals for the Fifth Circuit vacated a preliminary injunction issued by Judge Henry Wingate of the Southern District of Mississippi and ordered that the case go to a different judge. According to the appeals court, his orders contained AI hallucinations: parties who were never in the case, declarations missing from the record and a citation to a case that does not exist.

The published opinion in Jackson Federation of Teachers v. Fitch (No. 25-60496) is unusual: the consequence falls on a judge, not a lawyer.

What did the Fifth Circuit decide?

The Fifth Circuit vacated the injunction, meaning it no longer has effect, and directed the Chief Judge of the Southern District of Mississippi to assign the case "to a different district judge for all further proceedings".

Judge Jerry E. Smith wrote the opinion for a panel with Judges Haynes and Engelhardt. Judge Haynes concurred only in the final paragraph, which vacates, remands and orders reassignment, and stressed that teachers must remain free to teach students not to discriminate.

The underlying case concerns House Bill 1193, a Mississippi law of April 2025 that restricts diversity, equity and inclusion (DEI) programmes and trainings in public schools and universities. Teachers, students, parents and organisations challenged it in June 2025 before it was applied to anyone, arguing that it breaches the First Amendment and is too vague.

The ruling does not decide whether the law is constitutional. The panel held that the injunction rested on errors of law, mainly under the Supreme Court's decisions in Moody v. NetChoice (2024) and Trump v. CASA (2025), and set out the law the new judge must apply. This article takes no position on the dispute over the law itself.

How did AI errors get into a federal judge's orders?

According to the opinion, a temporary restraining order issued on 20 July 2025 contained statements found nowhere in the record. A temporary restraining order (TRO) is a short emergency order that freezes the situation until a fuller hearing. The original order:

  • named as plaintiffs eleven people and organisations that had never been parties;
  • named as defendants two officials who had not been sued in that capacity;
  • stated, without evidence, that three higher-education institutions had already cancelled DEI programmes because of the law;
  • attributed to the statute phrases such as "critical race theory", which do not appear in it;
  • relied on declarations from four people who are not in the record and, in the panel's words, "may not even exist".

On 22 July 2025 the state asked the court to clarify the order. The same day the court removed it from the docket and replaced it, without explanation, with a seemingly corrected version backdated to 20 July. It refused to restore the original to the record, calling the errors "clerical".

After US Senator Chuck Grassley accused the court of using AI, Judge Wingate acknowledged in an October 2025 letter that a law clerk had used AI to help draft an earlier version that was docketed by mistake. He said the errors did not concern "core legal reasoning or the ultimate judgment on the merits". The panel could not see how so many mistaken facts would not affect the reasoning at least marginally.

At oral argument it emerged that even the corrected order cited "Cousins v. School Bd. of City of Norfolk, 503 F.2d 422 (4th Cir. 1974)", a case the panel found does not exist. Because the preliminary injunction of 18 August 2025 incorporated the TRO's analysis by reference, the fake case arguably infected that order too.

Why did the appeals court take the case away from the judge?

The panel held that the judge would be expected to have substantial difficulty setting aside views built on false findings, and that an objective observer might question the impartiality of further proceedings before him.

US federal appeals courts may reassign a case on remand under their supervisory power and 28 U.S.C. § 2106, a power the panel called "extraordinary" and "rarely invoked". Fifth Circuit case law asks whether the judge would struggle to set aside views found erroneous (weighed with the appearance of justice and the cost of a new judge), and whether an objective observer might question the judge's impartiality.

Both tests were met. The judge had removed the hallucinated "facts" without explaining why the order still stood. Given "the embarrassment of the AI hallucinations and mistakes of law", the panel saw an understandable but unjust temptation for him to prove his first view right, and found that sending the case back to the same judge would diminish public esteem for the judiciary.

The panel called the proceedings "a botch job" and said the judge had issued "sloppy edicts based on falsehoods and fabrications". It imposed no sanction on him, however, and made no disciplinary referral; its remedy was to move the case.

Why does a judge's AI error matter more than a lawyer's?

A fabricated citation in a brief is a risk the court can catch; a fabricated citation in a judgment is the court itself speaking.

The leading cases have concerned lawyers, starting with Mata v. Avianca in 2023, as our analysis of two thousand fabricated citations shows. Cases about judges are rarer. In September 2026 an Arizona appeals court vacated a sentence because the judge had relied on an AI-generated video of the victim, and Colombia's Constitutional Court has set principles for judges' use of ChatGPT. The Fifth Circuit opinion adds two points:

  • The judge answers for the chambers. The opinion attributes the orders to the judge, whoever used the tool.
  • The record matters as much as the citation. Removing an order without explanation and backdating its replacement made review harder. The panel said it was "lucky" the defendants had filed a copy of the original as an exhibit.

Can judges use AI in Türkiye? What the rules say

We found no binding Turkish rule specifically on judges' use of generative AI, but existing rules already require a decision to rest on real facts and law, with reasons.

The closest guidance comes from the Council of Judges and Prosecutors (HSK), which appoints and disciplines judges. Its Guide to the Turkish Judicial Ethics Declaration, adopted by decision No. 2025/1269 of 15 October 2025, says that AI does not change the fact that judges and prosecutors are the main actors in judicial work, and that the legal process should remain under their personal control.

Article 141 of the Constitution requires every decision of every court to be written with reasons. In civil cases, Article 27 of the Code of Civil Procedure (HMK) makes concrete and clear reasoning part of the right to be heard. In criminal cases, Article 34 of the Code of Criminal Procedure (CMK) requires reasoned decisions, and under Article 289(1)(g) a judgment without the reasoning required by Article 230 is an absolute breach of law on appeal to the Court of Cassation.

Turkish law also limits silent corrections. Under HMK Article 304, a court may correct only "writing and calculation errors and other similar obvious errors", and after service it may not do so without hearing the parties. In my view, invented parties and declarations are not errors of that kind. A party who doubts a judge's impartiality can challenge the judge under HMK Article 36 or CMK Article 24.

There is better practice at home. In April 2026 the Istanbul 14th Commercial Court disclosed in its judgment that it had used AI to verify and translate foreign case law. In September 2026 the President of the Constitutional Court said that AI should support the judge, not replace him.

For lawyers the rules are sharper. Article 34 of the Attorneys Act requires care, accuracy and honour. The Union of Turkish Bar Associations' 2026 recommendation guide on AI use says nothing obtained from AI may be filed without checking it against an official, reliable source, and that a hallucinated precedent may be a serious breach of the duty of care. In a case reported in June 2026, a Kızılcahamam court notified the bar and prosecutors after a petition cited Court of Cassation decisions with AI-generated wrong case numbers.

How do Korea and Europe handle AI in judges' work?

Both have written guidance for courts, but nothing like the Fifth Circuit's remedy.

Korea. According to Herald Economy (헤럴드경제), the National Court Administration announced an "AI Guidebook for Judges" in February 2026 that covers hallucination and privacy and whose worked examples use de-identified data rather than raw case records. In March 2026, the paper reported, a court task force proposed measures against AI-fabricated filings, such as costs orders and referral of lawyers to the Korean Bar Association. In June 2026 the Patent Court recorded in its judgment that counsel had filed non-existent AI-generated precedents.

Europe. The Council of Europe's European Commission for the Efficiency of Justice (CEPEJ), whose members include Türkiye, adopted non-binding guidelines on generative AI for courts (CEPEJ(2025)18Final) on 4–5 December 2025. They call "legal hallucinations" a serious risk to legal certainty, keep ultimate responsibility with judges, who should examine every AI output, and give parties the right to know which parts of a decision were AI-assisted (paragraphs 52, 56 and 61). In the EU, the AI Act classes AI systems intended to help a judicial authority research and interpret facts and law as high-risk (Annex III, point 8(a)), and Recital 61 says final decision-making "must remain a human-driven activity".

What should courts and law firms take from the ruling?

AI use needs a record and a human check before anything is signed or filed. That means:

  • recording when AI is used to research or draft a decision or filing, and by whom;
  • checking every citation, quotation and fact against official databases and the case file;
  • correcting errors openly and keeping the original on the record;
  • disclosing AI assistance where it shaped a decision, as the Istanbul court did;
  • keeping case files out of public AI tools; as KVKK told lawyers, pasting a file into a chatbot may be a transfer of personal data abroad.

What to watch next

A new judge in Mississippi will apply the Moody framework as the panel described it; further review by the full Fifth Circuit remains possible under the ordinary rules. In Türkiye, watch whether the HSK or the Ministry of Justice issues practical rules on AI in judicial work, and how the Kızılcahamam referral ends.

Frequently asked questions

Did a federal judge use AI to write a court order?

Yes. According to the Fifth Circuit's opinion of 22 October 2026, Judge Henry Wingate of the Southern District of Mississippi admitted using AI to prepare a July 2025 temporary restraining order, blaming a law clerk's use of AI. The order named non-parties as plaintiffs and relied on declarations missing from the record.

Was Judge Wingate removed from the case?

Yes. In Jackson Federation of Teachers v. Fitch, the Fifth Circuit vacated Judge Wingate's preliminary injunction and directed the Chief Judge of the Southern District of Mississippi to assign the case to a different district judge for all further proceedings. The opinion imposes no sanction on the judge.

Did the Fifth Circuit uphold Mississippi's DEI law?

No. The Fifth Circuit vacated the preliminary injunction against House Bill 1193 and remanded the case, but did not rule that the law is constitutional. It set out the legal framework, including Moody v. NetChoice, that a new district judge must apply.

Can a judge use AI in Türkiye?

We found no Turkish statute that bans it and no binding rule written specifically for judges' use of generative AI. The HSK's 2025 Judicial Ethics Guide says the legal process should stay under the personal control of judges, and Article 141 of the Constitution requires every court decision to be reasoned.

Can lawyers in Türkiye write petitions with AI?

Yes, but the lawyer remains responsible. Article 34 of the Attorneys Act requires care and accuracy, and the Union of Turkish Bar Associations' 2026 guide says AI-generated case law or provisions must be confirmed from official, reliable sources before filing.

Burhan Doğuş Ayparlar's View

This case is less about AI than about candour and the record. A clerk's mistake with a tool is familiar. What led to reassignment was what followed: an order replaced without explanation, a backdated version, and the word "clerical" applied to invented parties and witnesses.

For Türkiye the lesson is concrete. A reason resting on a case that does not exist is no reason, contrary to Article 141 of the Constitution. The HSK Guide sets the right principle, but it does not tell a judge what to log, check or disclose. The Istanbul 14th Commercial Court has shown a workable model: say in the reasoning what AI was used for and what the judge decided alone.

Lawyers should read the decision as a mirror, not as relief: one verification standard should apply on both sides of the bench.

This article is for information only and does not constitute legal advice. It is based on the Fifth Circuit's published opinion in Jackson Federation of Teachers v. Fitch, No. 25-60496, filed on 22 October 2026 and consulted on 25 October 2026; facts about the district court proceedings and Judge Wingate's letter are taken from that opinion. Its footnote dates for the two letters do not match the sequence in its text, so no exact dates are given. Also consulted: the HSK Guide, CEPEJ(2025)18Final (French version; translations are the author's), Regulation (EU) 2024/1689 and the Turkish statutes cited (mevzuat.gov.tr). The Union of Turkish Bar Associations' guide was read in the PDF published on its website (barobirlik.org.tr); the Korean measures are as reported by Herald Economy; the linked decision notes partly rest on press reports. The analysis and assessments are the author's own.