Learning AI #33

In early May 2025, a video played in an Arizona courtroom. A man looked into the camera and spoke directly to the person who killed him.

The man was Chris Pelkey. He was 37 when Gabriel Horcasitas shot him during a road rage confrontation at a red light in Chandler, Arizona. Pelkey did not write those words. He did not record that video. He'd been dead for three and a half years.

His sister, Stacey Wales, built the video with her husband, who works in AI. They used a photo of Chris and an AI voice model trained on old recordings of him, then wrote a statement in his voice: what she believed he'd say to the man who killed him. Horcasitas had already been convicted of manslaughter. This was his sentencing hearing.

Judge Todd Lang handed down the maximum available sentence, 10.5 years, more than the nine years prosecutors had asked for. Stacey Wales had asked for the maximum too. Lang said so directly in his ruling: "You demanded the maximum sentence. And even though that's what you wanted, you allowed Chris to speak from his heart as you saw it."

He read the forgiveness in the video as proof of the family's character, not as an argument for going easier on Horcasitas. By some accounts, he also told the courtroom he loved the video.

Horcasitas's defense attorney, Jason Lamm, objected before the sentencing and has flagged the video as grounds for appeal. His argument: an AI recreation has no thoughts or feelings behind it, no matter how convincingly it mimics someone who did.

He also pointed to a contradiction. The AI version of Chris forgave his killer on camera. Trial testimony described the real Chris getting out of his car and challenging Horcasitas to fight, seconds before he was shot.

Gary Marchant, a law professor at Arizona State, called the statement "completely fake" and raised the question that matters more than this one case: if a family can build a version of a dead person that says whatever they want it to say, on the record, in front of a judge, what stops every sentencing hearing from having one?

The Arizona Supreme Court has since formed a steering committee to decide what AI is and isn't allowed to do in its courtrooms.

Nobody in this case lied. Stacey Wales believed she was giving her brother a voice. Judge Lang believed what he heard. None of it was verified, cross-examined, or verifiably Chris's own words. It was one family's best guess at grief, delivered in a hearing where that same family was asking the judge for the maximum sentence the whole time.

The tools that made this video are not expensive or hard to get. A photo, a few minutes of someone's voice, and a story you want told is enough to build a person who says exactly what you need said, in front of a judge, a jury, or anyone else you're trying to convince.

I previously wrote about AI voice cloning in this post. There were a few important takeaways. First, you can clone your voice with AI, and in ten minutes have a pretty good version and in an hour you will have a version that people cannot distinguish from your real voice. I included test audio recordings in that post and asked for readers to vote on which was real and which was AI-generated. The voting split 50-50, which means nobody could really tell.

Second, families should develop a safe word to use with children and elderly parents. The last thing you need is for the bad guys to clone your voice and start directing your kids to the corner to get picked up by a stranger.

I don't think Pelkey's case stays a one-off. The next family that wants their loved one to have the last word in court now has a template.

Things I think about

A day on Venus is longer than a year on Venus. It takes about 243 Earth days for Venus to rotate once, and about 225 Earth days to orbit the sun.

*****