The Victim Didn't Say That
This post is part of my Medium blog.
Last year, in a Phoenix courtroom, a dead man forgave the man who killed him.
He wore a grey baseball cap. He looked at Gabriel Horcasitas and said it was a shame they had met that day, in those circumstances. In another life, they probably could have been friends. He said he believed in forgiveness, and in a God who forgives.
Christopher Pelkey was shot at a red light in 2021. Every word in that courtroom was written by his sister, Stacey Wales, and spoken by a model. She built the video from his writings and posts, his photos, recordings of his voice. Her goal, she told Reuters, was "to humanize Chris, to reach the judge, and let him know his impact on this world and that he existed."
The judge was reached. "I loved that AI," Judge Todd Lang said from the bench. He told the family that as justifiably angry as they were, he had heard the forgiveness, and he felt it was genuine. Then he sentenced Horcasitas to ten and a half years for manslaughter.
Last week, the Arizona Court of Appeals threw the sentence out. (The New York Times has the ruling.)
One correction before it hardens into shorthand, because even the Times headline says "conviction overturned": the conviction is the one thing that survived. A jury found Horcasitas guilty, and that stands. What the three-judge panel vacated was the sentence. The video, they wrote, was "fundamental error": it "erases the interpretive distance between the family's belief of what the victim would have said and the victim's own voice and opinions." A prison term had rested on "the imaginings of the victim's sister." And on the question of whether the judge actually relied on it: "While the record does not reflect precisely how the AI video factored into the sentencing calculus, there is no question it played a role."
Here is what should have been impossible. "I loved that AI" is a sentence that should never come off a bench. A judge's job, when handed a moving artifact, is to ask what it is. Lang answered with an aesthetic. He loved it. It felt genuine. And that stood in for the evidentiary question nobody had asked. The video even vouched for itself: near the end, the recreated Pelkey says the video is "the true representation of who I was." A witness that swears itself in. No court allows that from a person, and no rule of evidence had been written with a machine in mind. That gap is why it happened at all.
And that is the strange part. This was not fraud. No one was scammed. A grieving family used the best new tool on earth to put their brother back in the room for a few minutes. A judge was honestly moved. The prosecutor defended the video. Everyone in that courtroom acted in good faith, and it was still a due process violation. Good faith, a new tool, and no rules: that is the complete recipe.
The legal system runs on testimony. Sworn statements by people who can be cross-examined, people who can be wrong, people who were there. Courts are one of the last institutions that still officially distinguishes between what happened and what we wish had happened. Generative models break exactly that distinction. A voice is no longer a person. An artifact with no one behind it walked into a sentencing hearing and was received as a man. Sixteen months later, an appellate panel wrote the first paragraph of the case law, and courts around the country began drafting rules about AI in evidence.
That arc, incident then opinion then rule, is what I built The Condition Set to capture. The trilogy is an attempt to watch how policy evolves as these systems arrive, one decision at a time, usually after the harm.
In Redundant, there are no rules at all. The tools arrive, headcount follows, and the only check on what the systems are doing is one FinOps director and his spreadsheets. The question the book keeps asking is not whether the numbers are right. It is who the numbers get used against. That is the world the Phoenix courtroom was living in: the tool existed, and nothing anywhere said what it was for.
In Contingent, certification and audit trails exist, and the failure is subtler. Nadia Cho, an orchestrator, has her credential used to authorize a decision she never saw, with her name on it and a timestamp. The log cannot tell her judgment from her signature. The AI Pelkey is that same shape of wrong: an identity borrowed to produce an utterance the person never made. The appellate court's phrase, "interpretive distance," is the borrowed-credential problem wearing a robe.
In Essential, set in 2033, the question has become law. Which decisions must remain human is being settled in public, vote by vote. Last week we learned one answer early: the victim's voice at sentencing. That one stays human. A model can wear his cap and his cadence, but forgiveness is not something his sister can delegate, and it is not something a machine can hold.
The books compress a decade of that policy evolution into three stories. The real world is doing it in real time, and mostly in prose like the Arizona panel's: careful, corrective, arriving after the fact.
The ruling will read as cold to Pelkey's family. It should not. What the court protected is the difference between Chris and our idea of Chris. His sister's belief about what he would say is evidence of her love. Her love is real. It is not evidence of his opinion. "In another life, we probably could have been friends" is a lovely thing for a man to say. It is not a thing a model can say for him.