A court reporter AI transcript problem has now made it into a judicial decision: Judge Paul Felix said a trial transcript in an appeal contained speaker-attribution errors that appeared consistent with generative AI assistance, and he warned that court records still have to be checked by the humans filing them.
Felix raised the issue in a July 23 memorandum decision in a case involving a man accused in connection with selling drugs to another man who overdosed and died. The warning appeared in a footnote, which attorney Rob Freund later flagged on X.
According to Felix, the transcript misidentified who was speaking at several points in the proceeding. A motion that appeared to have been made by the State was attributed to the trial court. An objection that appeared to have come from Williams was assigned to the bailiff. The State’s closing argument was also attributed to the trial court.
Felix wrote that those problems, along with others he did not list, made the appellate review harder, although they did not materially block the court from handling Williams’s appeal. He said the court depends on transcripts to be accurate accounts of what happened in the courtroom.
The judge did not say he had proof of which tool was used or name a transcription product. He wrote that, based on the kind of mistakes the court saw, generative artificial intelligence “may have assisted” in preparing the transcript. His warning was less subtle than a bench slap: AI may save time, but the person using it still has to proofread the output and make sure it is right.
What did the judge say about the AI transcript errors?
Felix said the errors looked like the kind of mistakes generative AI can produce when it guesses instead of resolving ambiguity. In a court transcript, confusing the judge, a lawyer, and the bailiff is not a cosmetic defect; it changes the official account of who said what.
That matters because appellate courts work from the record. If the transcript is wrong, judges reviewing a conviction or other ruling may have to spend time sorting out basic facts of the proceeding before reaching the legal issues. In this case, Felix said the errors complicated review but did not substantially prevent it.
The episode shifts the familiar courtroom AI problem from lawyers’ briefs to the record itself. Judges have repeatedly called out attorneys for submitting filings with AI-invented cases or unreliable citations. In May, judges in the Supreme Court of the State of New York Appellate Division criticized several lawyers after accusing one of them of using AI and others of failing to catch the problem.
Court reporting is a different failure mode. A bad legal citation can be checked against a database. A faulty transcript may become the database, unless someone catches it. That is why the distinction between draft assistance and official recordkeeping matters.
AI transcription tools are marketed to the legal industry, including court reporters. Stenographers have argued that trained human listeners still play a critical role because courtroom audio is messy, speakers interrupt, and ambiguity often has to be resolved before words are entered into the record. Felix’s footnote gives that argument a fresh exhibit.
This story draws on original reporting from 404 Media.
