A Court Can Correct an Order. It Cannot Undocket the Doubt.

The appellate court is asking a remedy question

The underlying case, Jackson Federation of Teachers v. Fitch, challenges provisions of a Mississippi law restricting diversity, equity, and inclusion programs in public education. In July 2025, U.S. District Judge Henry Wingate issued a temporary restraining order blocking parts of the law. State officials promptly filed an unopposed motion identifying errors that could be checked against the docket itself.

The order named plaintiffs who were not parties. It identified officials as defendants in capacities in which they had not been sued. It described allegations absent from the operative complaint, quoted terms that did not appear in the statute, and relied on declarations from four people whose declarations were not in the record. The court replaced the order and later characterized the problems as clerical errors arising from oversight or omission.

Months later, in a letter responding to Senate Judiciary Committee questions, Judge Wingate gave a more complete account. A law clerk had used Perplexity as a foundational drafting assistant to synthesize public docket information. The docketed document was an early draft that had not passed through the chambers' ordinary review process. The judge called its release a mistake and described new controls, including independent review by a second clerk and printed copies of cited cases attached to final drafts.

The appeal now before the Fifth Circuit concerns a later preliminary injunction. According to Reuters, the panel instructed counsel to address reassignment at the August 31 argument. That posture matters. The appellate court is not merely deciding whether the first order was embarrassing. It is asking what remedy, if any, is appropriate for proceedings that may continue after the appeal.

These were not ordinary citation errors

The legal profession has developed an unfortunate familiarity with hallucinated citations. A nonexistent case can often be exposed by searching a database, checking the reporter, or reading the opinion. The error is serious, but its shape is familiar: a proposition is attached to authority that does not exist or does not support it.

The Mississippi order presented a different category of failure. The draft did not merely misdescribe the law. It imported a different record. People who were not plaintiffs became plaintiffs. Declarations by people outside the docket became evidence. Language absent from the statute appeared as if quoted from it. Those errors affected the basic identity of the dispute the court was deciding.

That distinction should change the analysis. A false citation threatens the reasoning that connects law to outcome. A false record threatens the premise that the court heard this case at all. Judicial review cannot function if the order answers a synthetic version of the controversy assembled from material that the parties did not submit and could not contest.

The danger is not that an AI system secretly became the judge. The judge remained responsible for the order. The danger is more mundane and more plausible: a drafting system produced an authoritative-looking composite, and the process designed to distinguish the record from plausible context failed before the document acquired the force of law.

Correction fixes the text, not the history

Courts need the ability to correct orders. Typographical mistakes, mistaken dates, missing words, and accidental docket entries are inevitable in any human institution. A rule that treated every correction as evidence of incapacity would make judging impossible.

This is why transparency is more than public relations. A litigant evaluating a corrected decision needs to know whether the problem was isolated, whether the same process affected later work, whether the corrected order was rebuilt from the actual record, and what safeguards changed. Without that information, the litigant is asked to trust the result of a process the litigant cannot inspect.

Judge Wingate's letter supplied facts that were missing from the initial response and described concrete reforms. That deserves weight. It also illustrates the cost of delayed explanation. Confidence is easiest to preserve when the court promptly identifies what failed, what did not fail, and how the operative ruling was independently verified.

Reassignment is extraordinary and not necessarily punitive

Federal appellate courts possess supervisory authority under 28 U.S.C. § 2106 to direct further proceedings on remand as justice requires. The Fifth Circuit has repeatedly described reassignment as an extraordinary power that should be used rarely. It is not the appellate equivalent of moving a file because another judge might reach a different result.

The circuit has considered whether the original judge would have substantial difficulty putting aside prior views found erroneous, whether reassignment is advisable to preserve the appearance of justice, and whether the waste and duplication would be disproportionate to the gain in apparent fairness. In some cases, the court has also asked whether an objective observer might reasonably question the judge's impartiality. Personal bias is not the only possible concern.

That framework makes the Fifth Circuit's question more precise. Reassignment would not necessarily declare that Judge Wingate acted with bias, bad faith, or an improper substantive motive. It could instead reflect a judgment about institutional confidence in future proceedings. The remedy looks forward: can the same court continue without reasonable doubt about the integrity of the process, and is the cost of starting with another judge justified by the benefit?

The distinction matters because punitive language can obscure the legal test. Judicial discipline, recusal, reversal, and appellate reassignment are different mechanisms serving different purposes. The present appeal may end without reassignment. The fact that the panel raised the issue still signals that AI-related process failures can affect more than the validity of one document.

Judges and lawyers are not situated symmetrically

It is tempting to say judges should face the same rules imposed on lawyers. At a minimum, both must verify the authorities and facts for which they take responsibility. The Code of Conduct for United States Judges requires judicial duties to be performed fairly, impartially, and diligently, and requires judges to maintain professional competence in the law.

Yet the positions are not symmetrical. Opposing counsel can challenge a lawyer's filing before it becomes a judgment. A judge's order arrives with state authority already attached. It may restrain conduct, move money, change custody, delay a business, or determine which rights remain enforceable while the parties seek correction or appeal.

That makes internal judicial governance unusually important. The public cannot audit chambers, so chambers must maintain controls worthy of the authority their work receives. Tool logs, source restrictions, docket-only retrieval, citation verification, version history, and clear responsibility for final review are not bureaucratic extras. They are substitutes for scrutiny the adversarial system cannot ordinarily provide inside the courthouse.

The audit trail may matter more than the AI policy

A useful chambers policy should identify what sources a tool may access, whether nonpublic material may be entered, which tasks may be delegated, how citations and record references are checked, who approves a final version, and what evidence of review is preserved. If a draft can be docketed, the system should make the unreviewed state conspicuous and the publication step difficult to complete accidentally.

The best control may be provenance. Every factual sentence in a proposed order should be traceable to a docket item, transcript, stipulation, or fact properly subject to judicial notice. Every quoted statutory phrase should point to the enacted text. Every authority should resolve to an opinion that says what the draft claims it says. AI can assist with that work. It cannot be permitted to conceal whether the work occurred.

The same lesson applies outside courts. Any organization using AI for consequential decisions should be able to reconstruct which data entered the system, which human approved the result, and which controls actually ran. A policy stating that humans remain in the loop proves very little if no artifact shows what the human reviewed.

Judicial authorship is a process, not a typing test

Law clerks have always researched, drafted, challenged, and refined judicial opinions. No serious account of judging requires the judge to type every sentence. Authorship means something more demanding: the judge understands the record, owns the reasoning, decides the result, and ensures that the order speaks only with authority the court actually possesses.

Generative AI does not make that model obsolete. It tests whether the model was real. If a plausible draft can move from synthesis to docket without anyone noticing that its parties and evidence belong to another case, then the problem is not merely the tool. The institution had a signature without effective authorship.

The Fifth Circuit may decide that the corrected process and later proceedings make reassignment unnecessary. It may conclude that a new judge is required to preserve the appearance of justice. Until the court rules, both outcomes remain possible.

The larger principle is already visible. Courts can correct text. They can reverse judgments. They can replace procedures. What they cannot do by editing a docket entry is return the parties to the moment before an unreviewed synthetic record was announced as a judicial act. The remedy must account for the doubt that remains, not merely the words that were withdrawn.

This article analyzes a pending appeal and publicly available records for general educational purposes. The Fifth Circuit has requested argument on reassignment but has not decided the issue, found judicial bias, or resolved the merits of the underlying civil-rights dispute.