By Bill Hartzer · Published · How this site is researched
The other side of the bench
Most reports about artificial intelligence in litigation concern lawyers who filed briefs containing hallucinated citations (references to cases that do not exist). This part of the record concerns the other side of the bench: judges, law clerks, interns, tribunal members and court administrators. The matters are fewer, but each carries more weight. A lawyer's error sits in a filing that the court can reject. An error in a signed order becomes part of the decision itself, and a party who loses on the strength of it may have no way to test it except by appeal.
The entries here record what the court or chambers did with an AI system, how the error or the use came to light, and what a reviewing court, legislator or administrator did in response. They do not characterize what any court decided about AI generally.
Adopted proposed orders and chambers errors
Three patterns account for most of the record so far.
- A proposed order is adopted with a party's unverified citations intact. In Shahid v. Esaam the Court of Appeals of Georgia vacated a divorce-case order that relied on 2 fictitious cases and purported to have been prepared by the husband's attorney. In Payne v. State in 2026, the Supreme Court of Georgia vacated a murder defendant's new-trial order prepared by a prosecutor who acknowledged using AI to draft the State's briefs. A California appellate court reversed and reassigned a protective-order case in H.C. v. Contreras in May 2026 after the trial judge relied on a fictitious case drawn from counsel's brief.
- Chambers staff use a general-purpose system, and a draft reaches the docket before review. Judge Julien Neals of New Jersey withdrew an opinion in the CorMedix securities case after an intern's unauthorized use of ChatGPT. Judge Henry Wingate of Mississippi replaced a temporary restraining order after a law clerk used Perplexity to summarize the docket.
- A tribunal's own research produces false authority. In Pooja Ramesh Singh v. Jammu and Kashmir Bank, India's Supreme Court found that an insolvency tribunal had relied on 6 fabricated or misquoted precedents that no party had cited, and that the appellate tribunal had repeated them.
A fourth pattern is different in kind: judges who used a system deliberately and said so. Judge Kevin Newsom of the 11th Circuit set out his queries to ChatGPT, Gemini and Claude in 2 concurrences in 2024. A Cartagena labor judge disclosed his consultation of ChatGPT in a 2023 health-coverage ruling, which Colombia's Constitutional Court later reviewed.
Matters in this topic
8 matters, most recent first within each group. 7 were checked against the primary document, and the others rest on reporting and say so.
Orders containing errors
- Payne v. StatePrimary document readA Clayton County trial court denied a murder defendant's motion for new trial in an order prepared by the prosecutor, who later acknowledged using AI to draft briefs. The Supreme Court of Georgia vacated the order and suspended her from practice before it for 6 months.
- In re CorMedix Securities Litigation (Judge Neals)Primary document readJudge Julien Neals withdrew an opinion denying dismissal in the CorMedix securities case after defense counsel identified misstated case outcomes and quotations that did not exist. He later told the Senate that a law school intern had used ChatGPT without authorization.
- Jackson Federation of Teachers v. Fitch (Judge Wingate)Primary document readJudge Henry Wingate replaced a temporary restraining order that named parties not in the case and quoted language not in the statute. He later told the Senate a law clerk had used Perplexity. Mississippi asked the 5th Circuit to reassign the case.
- Shahid v. EsaamPrimary document readA Georgia trial court signed an order, apparently prepared by the husband's attorney, that relied on 2 fictitious cases. The Court of Appeals vacated the order, remanded for a new hearing and imposed a $2,500 penalty on that attorney.
Judges using AI openly
- Snell v. United Specialty Insurance (Judge Newsom)Primary document readIn Snell (May 2024) and Deleon (September 2024), 11th Circuit Judge Kevin Newsom wrote concurrences describing queries he put to large language models about the ordinary meaning of words. The majority decisions did not rest on those queries.
Oversight and inquiries
- Senate Judiciary inquiry into judges' AI usePrimary document readSenate Judiciary Chairman Chuck Grassley asked Judges Wingate and Neals whether AI was used in drafting error-filled orders. Both acknowledged staff use, and the Administrative Office of the U.S. Courts described its interim AI guidance.
Outside the United States
- Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd.Primary document locatedIndia's insolvency tribunal admitted a case relying on precedents that were fabricated or misquoted, found through its own research and repeated on appeal. The Supreme Court set aside both orders on July 2, 2026 and sent the case back.
- Colombia: Cartagena ChatGPT ruling and Sentencia T-323/24Primary document readA Cartagena labor judge disclosed that he consulted ChatGPT in a January 2023 ruling on health coverage for a child with autism. Colombia's Constitutional Court reviewed it in 2024, found no due process violation and ordered guidance for the judiciary.
Vacated orders, withdrawn opinions and oversight letters
Reviewing courts have generally acted on the order, not the judge. They vacate, remand, reverse or reassign, and in some cases instruct the lower court to write the replacement order itself. The Supreme Court of Georgia's instruction in Payne that the new order "shall not be prepared by counsel for either party" is the most specific instruction of this kind in this record. Sanctions in these matters have fallen on the lawyers whose filings supplied the false authority, not on judges.
Where the error came from chambers, the response has come from outside the case. In October 2025, Senate Judiciary Chairman Chuck Grassley asked Judges Neals and Wingate whether AI had been used and whether non-public information had been entered into it. Both acknowledged staff use and described new review procedures. The Administrative Office of the U.S. Courts described interim guidance, issued July 31, 2025, that cautions against delegating decision-making to AI and calls for independent verification of output. Colombia's Constitutional Court ordered its judicial council to issue guidance. India's Supreme Court called for zero tolerance for unverified AI-generated precedent from bar and bench alike. On appeal in the Mississippi case, the State has asked the 5th Circuit to reassign the matter to a different judge; that request was argued in August 2026 and is pending.
Signatures carry the same duty on the bench
Measured against the Expert Record Standard, most of these matters engaged Clause 3, independent authority. A citation, quotation or fact entered a signed document on a system's authority and was not confirmed against the source. Several also engaged Clause 2, non-delegation. In Payne the reviewing court required that the replacement order be the trial court's own work, and in Shahid the appellate court could not review reasoning that rested on authority that did not exist.
The chambers matters add Clause 7, contemporaneous record. In the New Jersey case the intern's use of ChatGPT was unauthorized and undisclosed, and in both federal cases the account had to be reconstructed months later in response to a Senate inquiry. The disclosed uses show the reverse. Judge Newsom's concurrences and the Colombian ruling could be examined precisely because the use was stated and the queries recorded, which is Clause 8, disclosure, in practice.
For an expert witness the transfer is direct. A report, like an order, is adopted by signature. Text supplied by counsel, by an assistant or by a system becomes the expert's own when signed, and every authority in it should have been opened and read before that point.
What errors on the bench do not show
The record does not show how often courts use AI without incident, or how many errors go unnoticed. The matters here came to light because a party, a lawyer or an appellate court checked. A 2026 survey report published by the New York City Bar Association found that more than 60% of responding federal judges had used at least one AI tool, with legal research the most common use. That figure describes use, not error.
Nor does the record show discipline of judges for AI-related errors in the matters recorded here, or a settled federal rule on disclosure of judicial AI use. In several matters the reviewing court expressly declined to find how the false citations got into the order. Where a court has not made that finding, this record does not make it either.
Whose document held the error
Reports of judicial AI errors travel fast and are often imprecise. Before relying on one, establish whose document contained the error: a party's brief, a proposed order the court signed, or text written in chambers. Establish who has confirmed that AI was used. A judge's own letter or an affidavit from counsel is a different kind of evidence from a commentator's inference drawn from the pattern of errors. Establish what the reviewing body actually did. Look for the order itself, or for the judge's or court's written account, and note the date, because withdrawn orders are often removed from public dockets and corrected versions may carry the original date.
Finally, check the status. Several of these matters were sent back for new orders or remain on appeal, and the outcome on remand is often reported, if at all, well after the original story.
Primary documents behind this topic
The orders, opinions and published texts that entries in this topic were checked against. Read them before relying on any summary, including the ones here.
- Payne v. State
Payne v. State, No. S26A0459 (Ga. May 5, 2026) - Senate Judiciary inquiry into judges' AI use
Letter from Robert J. Conrad Jr., Director, Administrative Office of the U.S. Courts, to Sen. Charles E. Grassley (Oct. 21, 2025) - In re CorMedix Securities Litigation (Judge Neals)
Letter from Judge Julien Xavier Neals to Sen. Charles E. Grassley (Oct. 20, 2025) - Jackson Federation of Teachers v. Fitch (Judge Wingate)
Letter from Judge Henry T. Wingate to Sen. Charles E. Grassley (Oct. 21, 2025) - Shahid v. Esaam
Shahid v. Esaam, No. A25A0196 (Ga. Ct. App. June 30, 2025) - Colombia: Cartagena ChatGPT ruling and Sentencia T-323/24
Corte Constitucional, Sentencia T-323 de 2024 (Aug. 2, 2024) - Snell v. United Specialty Insurance (Judge Newsom)
Snell v. United Specialty Insurance Co., No. 22-12581 (11th Cir. May 28, 2024)
Questions about this topic
Have judges been sanctioned for AI-generated errors in their orders?
No sanction of a judge has been reported in the matters recorded here. Reviewing courts have vacated or reversed the affected orders, and sanctions have been imposed on the lawyers whose filings supplied false citations. In the federal cases, the response came through a Senate inquiry and changes to chambers procedures.
Is a federal judge required to disclose using AI?
No binding federal rule was found requiring it as of September 2026. The interim guidance described by the Administrative Office of the U.S. Courts in October 2025 asks judges to consider whether AI use should be disclosed. Some individual judges have disclosed their use in published opinions.
Why do so many of these cases involve proposed orders?
Courts often ask a prevailing party to draft an order, and a signed proposed order carries the party's citations into the court's decision. In Shahid v. Esaam and Payne v. State, citations that originated with counsel appeared in signed orders. The Supreme Court of Georgia responded by directing that the replacement order in Payne not be prepared by counsel.
What does this mean for an expert witness who uses AI?
It means the same discipline applies to a report as to an order: whatever is signed is adopted. Every citation, quotation and figure should be confirmed against its source before signature. The use of the system should be logged as it happens and disclosed where rules or the engagement require it.
Did any court approve a judge's use of ChatGPT?
Colombia's Constitutional Court reviewed a Cartagena judge's 2023 use of ChatGPT and found no violation of due process, because the judge had reached his decision before consulting the system. It also found that transparency and accountability were not fully met, and it ordered guidance and training for the judiciary.