Tech experts warn against using AI clone of murdered woman's voice in court | KSL.com<br>MARKETPLACEClassifiedsCars<br>Log in

MARKETPLACEClassifiedsCars<br>Log in

Tech experts warn against using AI clone of murdered woman's voice in court

By Dave Cawley, KSL |<br>Posted - Aug. 10, 2026 at 9:05 p.m.

Save Story

KEY TAKEAWAYS

Lawyers oppose using AI-generated voice of murder victim Joyce Yost in court.<br>Experts warn AI voice may mislead jurors by mimicking human speech inaccurately.<br>Judge to decide on AI voice use by mid-Nov., with sentencing in Feb. 2027.

FARMINGTON — Lawyers for a man convicted in a decades-old capital murder case are hoping to block Weber County prosecutors from using an AI-generated simulation of the victim's voice at an upcoming sentencing hearing.<br>Douglas Lovell abducted and murdered Joyce Yost on the night of August 10, 1985 to prevent her from testifying against him in a rape trial. Yost's body has never been found, but Lovell admitted to the killing during plea negotiations back in 1993. He has twice been sentenced to death and twice had that sentence overturned on appeal.<br>Lovell is currently scheduled to be sentenced again by a jury next February. Weber County prosecutors previously asked a judge for permission to play AI-generated simulations of Yost's voice to the jury.<br>The prosecution created the Joyce Yost voice clone using an AI software tool from the company ElevenLabs. By providing the tool with audio recorded clips of Yost's actual voice, as well as a transcript of testimony Yost gave at a preliminary hearing prior to her disappearance, prosecutors were able to generate synthesized audio of her words.

In court, Lovell's defense team called two expert witnesses to push back against the idea of playing the AI voice clone before a jury.<br>The first, Sameer Patil, holds a PhD in computer science from the University of California Irvine and works as an associate professor of computer science at the University of Utah.<br>Patil summarized current research in the field of human-computer interaction, noting people have a tendency to treat AI-generated content as if it came from a real human.<br>"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."<br>Patil said this presents a risk that jurors might treat the AI reproduction of Joyce Yost's voice as if it were the real thing, even if they are told in advance the voice is artificial.<br>Patil also said because the text transcript of Yost's past testimony includes only her words and not any paralinguistic cues, such as inflection or cadence, it is impossible to determine how similar the AI model's output is to Yost's actual testimony.<br>"Sounding similar to Joyce Yost does not necessarily mean being accurate about it, and also absolutely does not mean being accurate about how she spoke those words at that testimony," Patil said. "That simply cannot be validated because the audio recording does not exist."<br>The defense team's second expert witness was Kenneth Marino, who holds a PhD in machine learning from Carnegie Mellon University. Marino works as an assistant professor at the University of Utah and is a former research scientist at Google DeepMind.<br>Marino described how generative artificial intelligence voice synthesis works, starting with training a text-to-speech tool using samples of existing text and audio.<br>"AI companies will generally try to use the data that's most easily publicly available," Marino said. "For audio, that's likely things like podcasts and YouTube videos."

Output from the tool can then be further refined, Marino said, using "weights." But he noted ElevenLabs does not publicly disclose its training data or the weights it uses to refine its AI models.<br>"So we essentially have no idea what data it was trained on," Marino said.<br>Lovell's defense has issued a subpoena to ElevenLabs, seeking those under-the-hood details about its generative AI voice tools. ElevenLabs has not yet responded to the subpoena.<br>The debate put the four-decade-old Joyce Yost case on the cutting edge of AI use in the courtroom. Typically, courts use human proxies to read transcripts live, in real time, when prior witnesses are unable to testify at subsequent hearings. Use of a homicide victim's prior sworn testimony, presented using an AI voice clone, has never been done in a potential death penalty case.<br>Judge Michael DiReda said he anticipated making a decision about whether to allow the AI voice clone by mid-November. Sentencing for Lovell is currently scheduled to begin in late February 2027. However, Lovell's defense team notified the judge they would likely request the sentencing be delayed.<br>The murder of Joyce Yost is the subject of season 2 of the KSL Podcasts series COLD.

&times;

The Key Takeaways for this article were generated with the...