Judge Catches Court Reporter Letting AI Corrupt the Official Trial Record

Judge Identifies AI-Generated Errors in Official Court Transcript

Judge Paul Felix of the Indiana Court of Appeals has identified what appear to be generative AI errors in an official court transcript, issuing what may be the first judicial finding that AI corrupted a trial record. 1

The case, Williams v. State, concerned a man convicted for selling drugs to another man who died of an overdose. In a footnote to a memorandum decision filed July 23, Felix catalogued errors that went well beyond misspellings. Witnesses' and attorneys' names were reported incorrectly. A motion by the State was attributed to the trial court. An objection was attributed to the bailiff. The State's closing argument was assigned to the trial court. 2

The opinion was joined by Chief Judge Tavitas and Judge Bradford. "These errors, among others not described herein, complicated but did not substantially impede our review of Williams's appeal," Felix wrote. 2

Then the judge identified the probable cause. "Based upon the types of errors reviewed, it appears that generative artificial intelligence may have assisted with the preparation of this transcript." 2

Attorney Rob Freund spotted the footnote on X, and Samantha Cole reported the decision at 404 Media. 1

A certified trial transcript is the record every appeal and every legal precedent is built on. When an appellate panel reviews a conviction, the transcript functions as ground truth: the words spoken, the objections raised, the arguments made. The record carries no caveat and no confidence score. It is presented as what happened in that courtroom. If the bailiff never made an objection but the transcript says he did, the appellate court inherits a fiction.

That structural vulnerability separates this case from incidents where lawyers submitted fabricated AI-generated citations in their own court filings. In those cases, an opposing counsel could scrutinize the filing, verify the citations, and challenge the fabrication before a judge ruled. The lawyer was the author, the filing was the artifact, and the adversarial process supplied a check. A transcript operates differently. It arrives certified as correct under Indiana Appellate Rule 28(B), which requires the court reporter to certify accuracy. 2 Downstream, judges, appellate panels, and future litigants cite it as established fact. The certification assumes a human produced the record. It was not designed to catch a machine that attributes closing arguments to the wrong speaker.

Felix was measured. The court acknowledged that "trial records are rarely if ever perfect," quoting a 2001 Indiana Supreme Court decision. 2 But his warning was pointed: AI can improve efficiency, but the professionals who use such systems must proofread and verify the output. 2

This may not be isolated. The same appellate court identified similar transcript errors in March 2026 in Orr v. State, though that opinion did not attribute them to AI. 2

The accountability gap Felix flagged has no obvious remedy in existing court rules. Indiana Appellate Rule 28(B) requires the court reporter to certify accuracy, but the rule does not distinguish between a transcript the reporter produced and one a machine generated that the reporter then stamped. When a machine fabricates dialogue or misattributes speakers in a certified transcript, those errors become part of the permanent legal record. Future cases that cite Williams, or any case built on a similarly corrupted transcript, inherit the damage without knowing it.

The appellate panel in Williams caught the problem. They noticed a bailiff making objections and a judge delivering closing arguments, and they traced the errors to their probable source. A different panel, reading a different transcript, may not have the same opportunity. By then, the transcript will be certified, filed, and treated as what happened in that courtroom, and the appeals built on that record will cite it as fact.