We spend a lot of energy worrying that artificial intelligence will hallucinate fake cases, or conjure quotes that the real cases cannot support, or lie about the facts in a way that even this Supreme Court would be unwilling to rubberstamp. The Damien Charlotin AI hallucination tracker is presently showing over 1600 cases impacted by AI mischief. It seems that no amount of public shaming has put a stop to lawyers smashing the send button on drafts with Beavis v. Butthead buried within a particularly imposing string cite.

But AI’s hallucinations can also prove revealing. The machine lacks the guile of a lawyer, allowing it to stumble into the right answer that the law works hard to suppress.

U.S. v. Smith, out of the Eastern District of Michigan, was flagged by eDiscovery Today’s Doug Austin. The defendant faces extremely serious charges involving the exploitation of children, which is worth stating plainly up front, because nobody is organizing a defense fund over what follows and the point has nothing to do with whether he presents a sympathetic case. The point is about whether he deserves to have a lawyer throughout the process.

During a post-arrest interview, the government produced an AI-generated transcript and handed it to the defense and the court. That transcript reflects the defendant stating “But I need to have an attorney here,” a significant statement since the authorities did not, in fact, provide him with an opportunity to access an attorney and kept the interview going anyway.

However, Judge Terrence Berg did not take the AI-generated text on faith, and listened to the original tape. In a footnote, he caught a key discrepancy:

“The transcript records Smith as saying ‘But I need to have an attorney here.’… This is not correct, however, because in the audio recording it is clear that Smith asked the question, ‘Do I need to have an attorney here?’”

The decision also notes, “This transcript was not reviewed or verified by a court reporter but was provided for convenience in following the audio recording.” Sort of underscores the enduring value of court reporters, huh? Even if AI takes over the transcribing process, an independent court reporter reviewing recordings for accuracy and the appropriate nuance could save judges and lawyers a lot of time.

Technology made an error, but the more consequential mistake is the state of the law. Courts have waged a decades-long war on Miranda, shifting the goalposts for the right to counsel to make it increasingly inaccessible to regular people. Davis v. United States, requires the invocation of the right to counsel to be “clear, unambiguous, and unequivocal.” In practice, this means no matter how much a defendant states they would like to have an attorney present, the courts will turn any utterance short of yelling “Accio advocatus!” while standing on one leg into an open door to law enforcement firing up the hot lamps.

Perhaps the most infamous account of the justice system’s newfound hostility to Miranda arose in Louisiana, where suspect asked for a “lawyer, dog” and the state’s high court ruled — in an 8-1 decision, mind you — that the request was too ambiguous to honor because, hypothetically, the cops may have assumed the man might have wanted a dog who had passed the bar. Which is ridiculous on its face, but also… so what if he did, he’s still asking for a lawyer! Air Bud had to do something after his basketball ended with that tragic ACL tear.

When an ordinary person, who is not a lawyer, under stress, in a room full of cops, refers to a lawyer in plain English, the courts have trained themselves to hear nothing.

And so it should come as little surprise that in this case, Judge Berg dutifully recites that, “simply mentioning the prospect of talking with an attorney, or waiting to talk until one is present, is not sufficient.” The Supreme Court built that standard in 1994 around a suspect who said “maybe I should talk to a lawyer” and got told that was too non-committal to prompt cops to stop the grilling. The three decades since have been one long project of whittling the right to counsel down further.

The transcript error is, definitely, on its face a cautionary tale about AI reliability. But from another angle, it reveals the broken state of defendants’ rights. Large language models are, at the end of the day, fancy prediction machines. They’re taking input and guessing at the most likely correct output. For good and ill, AI reflects the “common sense” consensus. And when it heard someone in a police interrogation ask about having an attorney, it understood that the most likely interpretation of the recording was that the person was asking to have an attorney.

None of which gets Smith anything. The law should recognize his request as invoking his right to counsel, but as of today it does not. The AI transcribed the interview incorrectly and the consequent ruling accurately reflects the state of the law. Police won’t stop if the defendant ever wavers from the most unnaturally stilted proclamation that “I will not answer any further questions without an attorney.”

But after three years of hand-wringing over AI making up fake cases, it’s worth remembering that technology is a neutral actor. It doesn’t come in with a conscious agenda — unless it’s Grok. So its obliviousness can cut both ways and deliver common sense answers that should embarrass the justice system.

But it won’t.

AI-Generated Transcript Mistake Doesn’t Impact Court Ruling: eDiscovery Case Law [eDiscovery Today]
U.S. v. Smith, No. 2:25-CR-20824-TGB-APP (E.D. Mich. June 15, 2026) [Minerva26]

Earlier: Suspect Asks For A ‘Lawyer, Dog,’ Willfully Ignorant Court Denies Comma, Counsel

Joe Patrice is a senior editor at Above the Law and co-host of Thinking Like A Lawyer. Feel free to email any tips, questions, or comments. Follow him on Twitter or Bluesky if you’re interested in law, politics, and a healthy dose of college sports news.

The post AI Error Nearly Gave Defendant His Constitutional Right To Counsel — Don’t Worry, A Human Judge Fixed That! appeared first on Above the Law.