What happened
It was reported that on 30 June 2025 Judge Julien Xavier Neals of the United States District Court for the District of New Jersey signed and docketed an opinion on a motion to dismiss in In re CorMedix Inc. Securities Litigation, No. 2:21-cv-14020. On 22 July the defendants checked the authorities and wrote to the court. They identified three cases whose outcomes had been stated backwards, with motions to dismiss described as denied when they had been granted, and passages attributed to decisions that did not contain them. The court withdrew the opinion after the defendants raised the errors.
It was reported that around the same period Judge Henry Wingate of the United States District Court for the Southern District of Mississippi withdrew a temporary restraining order docketed on 20 July 2025 in Jackson Federation of Teachers v. Lynn Fitch, No. 3:25-cv-00417. Reporting identified parties who were not involved in the case, misstatements of state law and a reference to a case that did not exist. In October both judges answered questions from the Senate Committee on the Judiciary. Neals said a law school intern had used ChatGPT for legal research without authorisation or disclosure, contrary to a chambers policy that had previously been communicated only verbally. He has since put that policy in writing for all law clerks and interns. Wingate said a law clerk had used Perplexity to assist with an early draft based on public docket information. The draft reached the docket before the standard review process. His chambers now requires a second law clerk to review each draft independently, with cited cases printed and attached to final drafts.
What an auditable version would have shown
In the New Jersey case the defendants found the problems by reading the opinion and checking its authorities. The public record therefore shows an external check working only after the document had been signed and docketed. An auditable version would preserve, during drafting and review, how AI was used, which authorities were checked independently, what source was used for verification and who completed the final review. That record would neither establish that an AI-assisted document was correct nor replace judicial review. It would show whether the required checks happened before the document acquired the authority of the court. The Senate correspondence eventually supplied that explanation, months after the orders were issued.
Where the gap was
It was reported that the errors in both orders could be checked against primary legal sources. A case exists or it does not, a motion was granted or denied, and attributed words appear in the judgment or they do not. In New Jersey lawyers for one of the parties performed those checks and identified the discrepancies. In Mississippi Wingate said an early draft bypassed the standard review process and reached the docket. A CitationVerifier compares each authority and attributed passage against a recognised legal database before release, while a ConductRecord preserves where AI assistance was used and who reviewed the result. Neither control replaces the person responsible for the document. They make verification and accountability visible within the process instead of leaving another party to establish them after the court has acted.
What governance should have looked like
It was reported that the Administrative Office of the United States Courts established an advisory Artificial Intelligence Task Force in early 2025 and distributed interim guidance to federal courts on 31 July 2025. The guidance warned against delegating core judicial functions to AI, recommended independent verification of AI-generated material and placed responsibility for AI-assisted work on the user. Both orders were already on the docket when the guidance arrived, the first by about a month and the second by eleven days. Neals subsequently converted a verbal restriction into a written policy. Wingate added a second reviewer to the drafting process. A document carrying judicial authority requires both, a rule understood by the people doing the work and a workflow step capable of catching errors before issue.
Failure Pattern: machine-drafted text entered a signed judicial order without anyone verifying its citations, and the error was found by a party reading the document rather than by any check inside the court.
Governance Principle: before a document that binds people is issued, every citation in it should be checked against the real source, and there should be a record of what was drafted with what assistance and who reviewed it.
The reference implementation of CitationVerifier and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- Letter from Judge Julien Xavier Neals to Senator Chuck Grassley on judiciary use of AI (US Senate Committee on the Judiciary)
- Letter from Judge Henry Wingate to Senator Chuck Grassley on judiciary use of AI (US Senate Committee on the Judiciary)
- Letter from the Administrative Office of the US Courts to Senator Chuck Grassley on judiciary use of AI
- Court Operations, Annual Report 2025 (Administrative Office of the United States Courts)
- Federal judges using AI filed court orders with false quotes, fake names (The Washington Post)