Tech experts warn against using AI clone of murdered…

In a case that brings decades-old litigation to the forefront of modern technology, Weber County prosecutors are seeking judicial permission to use an AI-generated voice simulation of murder victim Joyce Yost during a sentencing hearing. The defense team for Douglas Lovell, who was convicted of abducting and murdering Yost in 1985, is challenging the move, citing significant concerns regarding the accuracy and psychological impact of synthesized audio on a jury.

What happened

Prosecutors utilized an AI software tool from ElevenLabs to create the voice clone. By inputting audio clips of Yost’s actual voice alongside a transcript of her prior testimony from a 1985 preliminary hearing, the software synthesized an audio version of her words. This approach deviates from standard courtroom practice, which typically relies on human proxies reading transcripts when a witness is unavailable.

During court proceedings, the defense presented two expert witnesses to argue against the introduction of the recording. Sameer Patil, a computer science professor, testified that jurors might develop an emotional attachment to the AI output, treating it as a living entity despite knowing it is artificial. He further noted that because the original audio of that specific testimony does not exist, the AI cannot accurately replicate Yost’s original inflection, cadence, or paralinguistic cues, making the output impossible to validate for accuracy.

Machine learning expert Kenneth Marino highlighted the lack of transparency surrounding the AI model, noting that ElevenLabs does not publicly disclose its training data or the weights used to refine its models. The defense has issued a subpoena to ElevenLabs to obtain these internal details, though the company has not yet responded. Judge Michael DiReda is expected to rule on the admissibility of the AI voice clone by mid-November, ahead of the sentencing hearing currently scheduled for late February 2027.