Defense Experts Challenge AI Voice Clone in Death Penalty Case

Defense Experts Challenge AI Voice Clone in Death Penalty Case

A Utah murder case has become a testing ground for artificial intelligence in the courtroom, as defense attorneys challenge prosecutors' plan to use an AI-generated voice clone of a murder victim during sentencing.

Douglas Lovell faces his third sentencing hearing for the 1985 abduction and murder of Joyce Yost, whom he killed to prevent her from testifying against him in a rape trial. Weber County prosecutors want to present Yost's prior testimony to the jury using an AI voice clone created with ElevenLabs software, feeding the system audio samples of Yost's voice and a transcript of her preliminary hearing testimony.

The defense called two computer science experts to argue against the technology's use in court. Their testimony, first reported by KSL, raises fundamental questions about how jurors perceive AI-generated content.

Why it matters

This case represents the first time prosecutors have attempted to use AI voice synthesis of a homicide victim's testimony in a potential death penalty proceeding. The legal precedent could shape how courts nationwide handle AI-generated evidence, particularly in cases where the stakes include life and death decisions.

Experts warn of emotional attachment to AI

Sameer Patil, an associate professor of computer science at the University of Utah, testified that research in human-computer interaction shows people consistently treat AI-generated content as if it came from a real human, even when they know it's artificial.

"For example, ChatGPT. Everybody knows it's AI," Patil said. "There's no human involved but people are developing emotional attachments to it. People are cursing at it like they're mad. Or when they describe what it does, they describe it like it's a thing that's alive."

Patil argued this creates a risk that jurors might respond emotionally to the AI voice as if hearing Yost herself, despite being told the audio is synthesized. He also noted a critical limitation: the original transcript contains only words, not paralinguistic cues like inflection or cadence, making it impossible to verify whether the AI reproduction matches how Yost actually spoke during her testimony.

Black box problem in AI training

Kenneth Marino, a University of Utah assistant professor and former Google DeepMind research scientist, explained how generative AI voice tools are trained on publicly available data like podcasts and YouTube videos. He pointed out that ElevenLabs does not publicly disclose its training data or the weights used to refine its models.

"So we essentially have no idea what data it was trained on," Marino testified.

The defense has subpoenaed ElevenLabs for technical details about its voice synthesis system, but the company has not yet responded. Courts typically use human readers to present prior testimony when witnesses cannot appear at subsequent hearings.

Judge Michael DiReda said he expects to rule on whether to allow the AI voice clone by mid-November. Lovell's sentencing is scheduled for late February 2027, though the defense indicated they may seek a delay.

These details were first reported by KSL.