Executive Summary

Key Numbers

Source: reporting by FOX 10 Phoenix and ABC15 Arizona.

The case goes back to November 2021 in Chandler, Arizona. Two drivers collided near the intersection of Gilbert Road and Germann Road, and Christopher Pelkey, thirty-seven at the time, got out of his truck and walked toward Gabriel Paul Horcasitas's car. Pelkey was shot in the chest and died.

There were two trials. The first conviction was thrown out after it emerged that prosecutors had not turned over text messages from Pelkey's phone to the defense, and at the second trial Horcasitas was convicted of manslaughter and endangerment. In May 2025, Judge Todd Lang of Maricopa County Superior Court imposed ten years and six months.

A year and four months later, on September 30, 2026, the Arizona Court of Appeals, Division One, vacated that sentence. The case number is 1 CA-CR 25-0191. The manslaughter conviction was left untouched, and only the sentence was lifted out and returned to Maricopa County Superior Court. What put Horcasitas back in front of a sentencing judge was a single video played in the courtroom.

The video was made by Pelkey's sister, Stacey Wales. The idea came to her while she was collecting written victim statements from family and friends ahead of the sentencing hearing, forty-nine of them in all, and she finished it with help from her husband and a friend who had worked with AI tools. The material came in two kinds. On one side were the videos and photographs taken while Pelkey was alive, and his voice. On the other were the script Wales wrote and the AI-generated passages that read that script aloud.

In the finished video, Pelkey talks about his faith, his family, and his forgiveness of the man who shot him. The longest single stretch is the part about forgiveness, and that stretch is what the trial judge held onto when he chose a sentence.

"In another life, we probably could've been friends. I believe in forgiveness and in God who forgives."

From the AI-generated video of Christopher Pelkey, played at the May 2025 sentencing hearing

Pelkey never said this. He died in 2021 and left no word anywhere that he forgave Horcasitas. The sentences were written by his sister, and what turned them into her brother's voice and face was AI. So the video holds two layers of different character. The lower one is a record of things that happened. The upper one is what somebody close to him imagined he felt. The line between them was plainly there while the video was being assembled, and it did not survive into the thing that played.

When the video ended, Judge Todd Lang said he loved that AI. It was a reaction to the screen he had just watched, and he moved that reaction over into the reasons for a sentence.

"I felt like that was genuine; that his obvious forgiveness of Mr. Horcasitas reflects the character I heard about today."

Judge Todd Lang, Maricopa County Superior Court, May 2025

The judge read the forgiveness on the screen as evidence of who Pelkey was, and imposed ten years and six months instead of the nine years and six months prosecutors sought. The reading should not tilt all one way. ABC15 reported that ten years and six months is the presumptive term for this offense. The judge did not step outside the range the law sets, and the length of the sentence was not what the appeals court took issue with.

That the video was made with AI was never hidden in the courtroom. The family said so, and the judge called it AI when he spoke. Everyone knew what it was made of, and the judge still took the forgiveness inside it as the dead man's own.

On appeal, Horcasitas's side challenged the reliability of the video and argued that the judge's reliance on it violated due process. Prosecutors answered that Wales had been close enough to her brother to know what he would have said. Division One did not accept that. A victim has the right to speak at a sentencing hearing, the court began, but that right does not override a defendant's right to due process.

In the opinion the court wrote down not what the video added but what it took away. The whole decision to vacate sits inside one sentence.

The video "erases the interpretive distance between the family's belief of what the victim would have said and the victim's own voice and opinions," presenting "a depiction of the victim and his thoughts created from the imaginings of the victim's sister."

Arizona Court of Appeals, Division One, September 30, 2026 (case no. 1 CA-CR 25-0191)

Interpretive distance is the center of this ruling. When a family member stands up in court and says that her brother would have put it this way, everyone listening knows they are hearing the family's reading. A gap stays open between the reading and the fact, and a judge adjusts the weight while looking at that gap. The video filled the gap in. The speaker wears the victim's face and carries his voice, so the words land as testimony rather than as interpretation. That, the appeals court held, is what made the sentencing fundamentally unfair.

The most practical part of the ruling is that the court did not throw the whole video out. The real footage of Pelkey, recorded while he was alive and mixed into the video, may come into a sentencing hearing; the passages AI had generated may not. The line the court drew falls on provenance rather than on medium. What was actually recorded is allowed, and words never recorded, dressed up to look like a record, are not.

Brownstone Law, a legal news outlet, added one point in its summary of the ruling. Disclosing that the video had been artificially made did not resolve the problem. Disclosure and separation are different jobs. Saying that a thing was made is one line attached to a whole file, and telling a recorded passage from an invented one happens inside it.

Carry this case out of the courtroom and the shape turns familiar. Authentic material exists, a human reading goes on top of it, and the two merge into a single output. A summary that goes into a report, an estimate laid over customer data, synthetic rows filling the space between measured ones all share that structure. Whoever builds it knows which part is which. But that knowledge does not ride out with the output.

The Pelkey video showed exactly that point. At the production stage, the real footage and the sister's script were plainly separate files. After rendering, it was one video of one person speaking without a break. The boundary disappeared inside the result, and the receiving side had no way to restore it. A courtroom is the strictest place there is for examining evidence, and the person sitting there was a judge. It was still not restored.

Lay the process out from left to right and the box where the appeals court drew its line becomes visible. The line sits between the two boxes on the far left, and the further right it travels, the more thoroughly that line is erased from the output.

The video arrived with an explanation that it had been made with AI, and that explanation by itself prevented nothing. A file-level label says only that something inside was generated. What the person judging needs is the next thing down. Which sentence, which figure, which frame is a record, and which one is generated. Without that resolution a label is only a warning, and the choice left is to believe the whole thing or discard it.

For an organization that works with data, this ruling leaves two questions. First, does our material mark the original passages and the generated passages separately? That asks whether the material itself records a sentence in an analysis report as a measurement or a model estimate, and a row of a training set as collected or synthesized. Second, does that marking travel downstream? If the distinction attached to the original falls away as the material is processed, summarized and redistributed, the person reading it at the end stands where Judge Todd Lang stood.

When AI-Ready Data comes up, format and quality are usually what come to mind first. This ruling adds a line to that. For data to be read by a machine, the values are not enough on their own, and where each value came from has to travel with it and carry the same standing. Material without that record is not material you can use. It is material that only lets you choose whether to believe it.

Thank you for reading this far. The reporting on this case is at AZFamily, ABC15 and FOX 10 Phoenix. Take a look at how far the boundary between original and generated survives in the material your own organization sends out, and whether that boundary reaches the last reader. If you find a place where it stops, we would like to hear about it.

• 1.AZFamily Digital News Staff. (2026). "Court of Appeals vacates Chandler road rage sentence due to AI victim video." AZFamily.
• 2.ABC15 Arizona. (2026). "Arizona appeals court throws out sentence after judge relied on AI-generated victim video."
• 3.FOX 10 Phoenix. (2026). "Arizona manslaughter sentencing vacated due to use of AI 'victim impact statement'."
• 4.Brownstone Law. (2026). "Arizona Court Vacates Sentence Over AI Victim Video."