Let’s talk about the dead man a courtroom brought back to speak — in words he never got to choose.
On the last day of September, the Arizona Court of Appeals threw out a sentence. Not the conviction. The sentence.
Gabriel Horcasitas, 55, was found guilty of manslaughter for shooting Christopher Pelkey, 37, at a red light in a Phoenix suburb back in 2021. Road rage. A confrontation most people have had at least once, except this one ended with a man dead. A jury rejected self-defence and convicted him. He was sentenced to ten and a half years.
Then, in May 2025, at that sentencing hearing, the dead man stood up and spoke.
Except he didn’t. Chris Pelkey had been dead for four years. What the courtroom saw was a video. It wore his face. It carried his voice. And what it said, in his mouth — his forgiving the man who shot him, addressing Horcasitas by name — no one had ever heard him say. Not once. Because he never got the chance.
Now, I want to be careful with this, because the easy version of the story is “AI faked evidence.” That’s not what happened, and if that’s the lesson you take, you’ve already missed the interesting part.
Nobody invented the family’s grief. Nobody fabricated the forgiveness. Pelkey’s family said the same thing over and over — through nearly fifty letters from friends and relatives, through everything they put on the record — that this was a man who would have forgiven. That belief was real. It was human. It’s the kind of thing courts are built to hear from families.
What was synthetic was the man delivering it.
And that distinction is the whole case. The appeals court, in an opinion written by Judge D. Steven Williams, put its finger on exactly one thing. The video, the court said, “erases the interpretive distance between the family’s belief of what the victim would have said and the victim’s own voice and opinions.”
It presented, in the court’s words, “a depiction of the victim and his thoughts created from the imaginings of the victim’s sister.”
Sit with that for a second. The imaginings of the victim’s sister. That’s not a cold swipe at a grieving woman. That is the court drawing a line and being precise about where it sits. Because there is a real difference between a family standing up and saying, “we believe he would have forgiven you,” and that same family’s grief putting on his face and his voice and delivering the verdict itself.
Distance. That’s the word. The court even named it. Interpretive distance.
And here’s why it matters so much more than the headline suggests. The court did not ban AI from the courtroom. It didn’t rule that synthetic media can never come near a sentencing hearing. Go read the opinions; that’s not in there. What it said was that this particular use crossed a line, that the judge expressly relied on it, that “no question it played a role,” and that the effect was “so prejudiced Horcasitas as to render the sentencing procedure fundamentally unfair.” The conviction stands. The sentence gets re-run.
Because the problem was never that the machine lied. The problem is that the machine collapsed a gap that the whole idea of testimony depends on.
Think about what testimony actually is. It’s not just information. It’s a specific person, accountable for specific words, standing in a specific moment where those words can cost them something. That’s why witnesses get cross-examined. That’s why the law cares who said it, not just what was said. The credibility is the person. Strip the person out and dress the statement in their skin, and you’ve produced the sound of an accusation without an accuser.
And there’s a detail from the reporting that ought to bother you more than the ruling. The synthetic Pelkey, in the video, told the court that it was the “true representation of who I was.”
A machine stood in a courtroom and testified to its own authenticity. In the dead man’s voice.
You can’t cross-examine that. You can’t ask it a follow-up. You can’t catch it in a lie, because it isn’t lying — every word in it may well be honest. It’s just standing in a place where no human can stand, using the one thing the human can no longer do anything about: their absence.
So here is what people are missing.
First. The fix here was not detection. Nobody in that courtroom needed a better deepfake detector to figure this out. The AI video was disclosed. Everyone knew it was AI. There was, according to coverage of the opinion, even a disclaimer attached — and the panel found the disclaimer didn’t cure the reliability problem. Because the issue isn’t whether the audience knows it’s synthetic. It’s whether the record keeps the seam between who said it and who imagined it. Provenance beats detection. Every single time.
Second. Nobody asked the dead man. In a room built entirely around the consent and accountability of whoever is speaking, the one person with the strongest claim on those words — the man whose face it wore and whose voice it borrowed — is the one person who couldn’t object. And that should be uncomfortable, because it’s not an edge case. It’s the entire coming decade. We have the technology now to make anyone say anything, and the first place it lands is the place where the dead can’t refuse.
Third, and this is the operator one, the lesson isn’t “don’t build likeness.” It’s: keep the seam. If you synthesize a voice, a face, a letter, a quote, a testimonial — for a brand, for a memorial, for a client, for a campaign — keep the human source and the machine’s rendering as visibly separate layers. Whose words. Whose voice. Whose belief. The moment you stack those into one seamless human-looking thing, you are lying, even if every phrase inside it is true. The distance is the integrity. The instant you erase it, you’ve done the thing this court just spent a whole opinion trying to name.
Because the machine can give a dead man one more sentence. It can give a father one more piece of advice, an artist one more song, a victim one more word of forgiveness. Some of that will be a genuine mercy, and I’m not going to pretend otherwise — the technology will comfort people, and that’s not nothing.
But it can never give him the right to have said it. Or the right to have stayed silent. And that right — the right to be quoted instead of rendered, to be the author of your own mouth and not a puppet of someone’s best guess — is the last thing that keeps a courtroom, or a family, or a whole internet, from becoming a room full of voices with no one left in them.
Let’s not trade that away just because the machine can make the ghost sound more convincing than the person ever did.
You’ve got this.