Arizona Appeals Court Vacates Road-Rage Sentence Over AI Video of the Dead Victim

TL;DR

On September 30, a three-judge panel of the Arizona Court of Appeals, Division One vacated the 10.5-year manslaughter sentence of Gabriel Paul Horcasitas, who shot and killed Christopher Pelkey in a 2021 road-rage confrontation. The reason: at sentencing, the judge watched and openly praised an AI-generated video in which a recreated Pelkey spoke to the court and forgave his killer. The conviction stands, but Horcasitas gets a new sentencing hearing. The court's core line is one every builder of likeness and avatar tools should read twice: the AI video "erases the interpretive distance" between what a family believes someone would say and that person's own voice, and "no disclaimer could cure that error."

What happened in the courtroom

Pelkey, a 37-year-old Army veteran, was killed on November 13, 2021, after he got out of his truck at a red light and walked toward Horcasitas's car, according to the Associated Press. After a first conviction was thrown out over undisclosed phone records, a second jury convicted Horcasitas of manslaughter.

At the May 2025 sentencing, friends and family gave victim impact statements, and Pelkey's brother read a statement Pelkey had actually written about his beliefs. Then his sister, Stacey Wales, played a video. She wrote the script; her husband Tim and a friend who had experience building humanlike AI avatars produced it from a single edited photo of Pelkey plus his voice and speech patterns pulled from existing footage, per the AP.

The opinion describes the video's structure in detail:

• The AI Pelkey opens by explaining it is a "version" of him "recreated through AI" using his picture and voice profile.
• A real clip of Pelkey talking about his faith plays.
• The AI Pelkey returns and says the video "is a true representation of who I was, you know, not how the courtroom portrayed me."
• It thanks the court and his family, tells Horcasitas "I believe in forgiveness and in God who forgives," and signs off over what appears to be real footage of him fishing.

"I loved that AI"

The problem was not that the video was shown. It was what Judge Todd Lang said afterward. From the transcript quoted in the opinion:

I loved that AI. Thank you for that... I feel that that was genuine; that his obvious forgiveness of Mr. Horcasitas reflects the character I heard about today... you allowed [Chris] to speak from his heart as you saw it.

He then imposed the presumptive 10.5-year term, which, because the State alleged no aggravating factors, was the maximum available. That was more than the prosecutor asked for. The avatar forgave Horcasitas; the sentence did not, which is the kind of irony defense lawyers frame and hang on the wall.

Why the court said it crossed the line

Arizona's Victims' Bill of Rights gives victims a broad right to be heard at sentencing, unsworn and without cross-examination. Sentencing also uses looser evidence rules than trial. But due process still requires that anything the judge relies on carry "minimal indicia of reliability," and the panel found the AI video did not.

Horcasitas's lawyers had not objected at the hearing, so the court reviewed only for fundamental error, a high bar. It cleared it anyway. Presiding Judge D. Steven Williams, joined by Judge Anni Hill Foster and Vice Chief Judge David D. Weinzweig, compared the video to the closest Arizona precedent, a death-penalty case where photos of a victim's young sons at his grave came "uncomfortably close" to the line but were allowed. The difference, per the opinion:

Unlike the photographs in Rose, which captured a particularly poignant moment for the young surviving victims, the AI video here (excluding the embedded "real" videos of the victim) does not reflect actual events... the AI video presents a depiction of the victim and his thoughts created from the imaginings of the victim's sister.

And then the passage that matters well beyond Arizona courtrooms:

It erases the interpretive distance between the family's belief about what the victim would have said and the victim's own voice and opinions. Indeed, the AI victim stated, "this here is a true representation of who I was, you know, not how the courtroom portrayed me." No disclaimer could cure that error.

Think of the difference between a eulogy and a ventriloquist act. A sister saying "Chris would have forgiven you" is obviously her reading of him. The same words in his face and voice, with his micro-expressions, arrive as testimony from the dead, and the label at the start does about as much as the "dramatization" caption on a true-crime reenactment.

The opinion is published, so it is precedent for Arizona trial courts. The panel noted that no Arizona case had ever addressed an AI-generated depiction of a victim, which tracks: when AP covered the original hearing, the video was believed to be a first in U.S. courts.

Reactions

Wales told ABC15 the goal was "to have Chris give an opinion of what we thought his opinion would be," and that "if removing the layer of AI erases any doubt what the sentence will be, that the judge decides, I'm all for it." Defense attorney Jason Lamm was blunter: "I appreciate the fact it was a cathartic moment for the victim's family, but the video was cringe." Lamm told ABC15 the new sentence could land between 7 and 10.5 years and that the case could still reach the Arizona Supreme Court. No resentencing date had been set.

What this means if you build with likeness

This is a criminal-procedure ruling, not a product regulation. But it is the clearest judicial statement yet of where synthetic first-person media breaks trust, and it lands on exactly the tools people are shipping: grief avatars, "talk to your late parent" apps, digital-twin video generators, voice clones.

Labels are not the fix

The video disclosed it was AI in its first seconds. The court said that did not matter, because the format itself delivers imagined words as the person's own. If your product's safety story is "we watermark it," this opinion is a reminder that a disclosure and an impression are different things.

First person is the risky voice

The court had no problem with the real footage spliced into the same file. The error was the synthetic speech presented as Pelkey talking. A tribute that narrates about someone, or that sticks to their recorded words, sits on very different ground than one that speaks as them.

High-stakes contexts will get gatekept

The court said it reviews victim-impact media case by case and that judges should screen presentations in advance. Expect lawyers, courts, and eventually insurers and HR departments to start asking where an avatar's words came from before letting it into a room where decisions get made.

Key Takeaways

• Arizona's Court of Appeals vacated Gabriel Horcasitas's 10.5-year manslaughter sentence on September 30 and ordered resentencing; the conviction stands.
• The trigger was a family-made AI video of victim Chris Pelkey forgiving his killer, which the sentencing judge said he "loved" and found "genuine."
• The court held the synthetic portions were unreliable because they depicted imagined thoughts, not real events, while the embedded real footage was permissible.
• "No disclaimer could cure that error": an upfront AI label did not save a video that presented imagined words as the victim's own.
• The published opinion is Arizona precedent and the first Arizona appellate ruling on an AI-recreated victim at sentencing; the defense says the case may still go to the state Supreme Court.
• For anyone building likeness, avatar, or grief tech, the line the court drew is first-person synthetic speech, not AI involvement as such.

Sources: State v. Horcasitas, 1 CA-CR 25-0191 (Ariz. App. Sept. 30, 2026), opinion PDF, ABC15, ABA Journal, Associated Press (May 2025)