A four-decade-old murder case in Utah is about to test uncharted legal ground: whether a jury can hear a synthetic replica of a dead woman’s voice, built from an AI voice clone using ElevenLabs software, as evidence in a death penalty sentencing. Weber County prosecutors say the technology could let jurors hear Joyce Yost’s own words again. Defense attorneys and independent tech experts say it could do something far more dangerous — mislead the very people deciding whether a man lives or dies.
Summary
Key takeaways
- Weber County prosecutors want to play an AI-generated voice clone of murder victim Joyce Yost during Douglas Lovell’s upcoming resentencing hearing.
- Yost was murdered in 1985 by Lovell to stop her from testifying against him in a rape case; her body was never found.
- The clone was built with ElevenLabs software using audio clips and a transcript of Yost’s preliminary hearing testimony.
- Two University of Utah experts warned the AI voice could unfairly sway jurors and cannot be verified as accurate.
- Judge Michael DiReda is expected to rule on admissibility by mid-November, with sentencing currently set for late February 2027.
Prosecutors Propose Using AI Voice Clone of Murder Victim
Prosecutors in Farmington, Utah, have asked a judge for permission to play a synthetic reconstruction of Joyce Yost’s voice in front of the jury that will decide Douglas Lovell’s fate. The request centers on a piece of testimony Yost gave decades ago, before she was killed and could never take the stand again.
Douglas Lovell abducted and murdered Yost on the night of August 10, 1985, prosecutors say, to keep her from testifying against him in a separate rape trial. Yost’s body has never been surfaced, but Lovell admitted to the killing during plea negotiations in 1993. He has been sentenced to death twice, and both sentences were later overturned on appeal. A jury is now scheduled to sentence him again next February.
Creation of AI voice clone using ElevenLabs software
To build the reconstruction, the prosecution’s team turned to ElevenLabs, an AI voice-synthesis company, feeding its software real audio clips of Yost speaking along with a written transcript of testimony she gave at a preliminary hearing shortly before she disappeared. The combination let the tool generate new synthesized audio meant to sound like Yost delivering those same words. It marks one of the first known attempts to use a cloned voice via AI as courtroom evidence tied to a homicide victim.
Defense Experts Highlight Risks and Limitations of AI Voice Evidence
Two computer scientists called by Lovell’s defense team argue that an ElevenLabs AI voice reconstruction, however convincing, cannot be trusted the way a live witness or an actual recording could — and that jurors may not be equipped to tell the difference.
Sameer Patil, an associate professor of computer science at the University of Utah who holds a PhD from the University of California Irvine, told the court that people routinely form emotional attachments to AI-generated content even when they know it isn’t human. “Everybody knows it’s AI,” Patil said, referring to tools like ChatGPT. “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.” He warned jurors could respond to the synthetic Yost voice the same way — treating it as real testimony rather than a machine-generated approximation, even after being told otherwise.
Patil also raised a more technical concern: the transcript used to build the clone captures only Yost’s words, not how she actually said them. Without inflection, cadence, or tone — what researchers call paralinguistic cues — there’s no way to check whether the AI output matches how Yost really sounded on the stand. “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.”
Concerns about undisclosed training data by ElevenLabs
The second defense witness, Kenneth Marino, an assistant professor of machine learning at the University of Utah and a former research scientist at Google DeepMind, walked the court through how text-to-speech systems are typically built. AI firms, he said, tend to train their models on whatever audio is easiest to find in bulk online. “For audio, that’s likely things like podcasts and YouTube videos,” Marino said. Output can then be fine-tuned using internal parameters known as weights — but ElevenLabs, Marino noted, does not publicly disclose either its training data or those weights. “So we essentially have no idea what data it was trained on,” he said.
That opacity has become a legal sticking point. Lovell’s defense team issued a subpoena to ElevenLabs seeking details about how the voice-cloning tool was trained and refined. As of now, the company has not responded, leaving the court without independent verification of how the synthetic Yost audio was actually generated.
Unprecedented Legal Questions Surround AI Use in Capital Murder Sentencing
Courts have long allowed human proxies to read a deceased or unavailable witness’s transcript aloud when that person can’t testify again. What hasn’t happened before, according to the case record, is presenting a homicide victim’s sworn testimony through an AI-generated voice clone in a case where the death penalty is on the table. That distinction is why this dispute over an ElevenLabs AI voice reconstruction is drawing attention well beyond Weber County — it could set a template, or a warning sign, for how courts elsewhere handle synthetic evidence.
Judge Michael DiReda has said he expects to rule on whether the AI voice clone can be played for jurors by mid-November. Sentencing for Lovell is currently scheduled to begin in late February 2027, though his defense team has already signaled it will likely ask for a delay — meaning the timeline, much like the underlying legal question, remains unsettled.
Why this matters beyond one Utah courtroom: as generative voice tools become cheap and widely accessible, prosecutors and defense attorneys alike will face growing pressure to decide when a synthetic reconstruction crosses the line from helpful illustration to unreliable — or unfairly persuasive — evidence. A ruling in this case could become one of the first real-world benchmarks for how American courts treat AI-generated speech in matters of life and death.
FAQ
Why do prosecutors want to use an AI-generated voice clone of Joyce Yost?
Prosecutors aim to use the AI voice clone to play synthesized audio of Yost’s past testimony at the upcoming sentencing hearing, arguing it lets the jury hear her own words even though she cannot testify herself.
What concerns have experts raised about using the AI voice clone in court?
Experts warn that AI voice cloning may emotionally bias jurors and lacks the paralinguistic cues, such as tone and cadence, needed to verify its accuracy against the original testimony.
Has AI-generated voice evidence been used before in death penalty cases?
No. Using an AI voice clone of a homicide victim’s testimony in a death penalty case is unprecedented, making this dispute a potential first-of-its-kind legal precedent.
When will the judge decide whether the AI voice evidence is admissible?
Judge Michael DiReda is expected to make a ruling by mid-November, ahead of Lovell’s sentencing hearing currently scheduled for late February 2027.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
