Flarepoint

State v Horcasitas: Sentence Vacated Over AI Victim Video

2026-10-07

On 30 September 2026, the Arizona Court of Appeals vacated the manslaughter sentence of Gabriel Paul Horcasitas because the sentencing judge had relied on an AI-generated video of the man he killed, Christopher Pelkey, appearing to speak in court. The court held the video was not reliable: it did not document anything Pelkey actually said or did, and a disclosure that AI had been used did not fix that. It is the first Arizona case to address an AI-generated depiction of a victim, and it draws a clear line between material that records an event and material an AI system generates — a line that matters to anyone who has to put AI output in front of a court.

What happened in State v Horcasitas?

Pelkey, 37, was shot during a road rage encounter at a stoplight in Chandler, Arizona, in November 2021. At a second trial, a jury rejected Horcasitas's claim of self-defence and convicted him of manslaughter.

At the sentencing hearing in May 2025, after relatives and friends had spoken, the family played a video that Pelkey's sister, Stacey Wales, had made using AI to recreate his likeness and voice. According to the appeals court's opinion, the video opened by explaining it was a version of Pelkey "recreated through AI" from his picture and voice profile. The same opinion records that the video included real footage of him, and that the AI figure told the court it was "a true representation of who I was" before addressing Horcasitas and speaking of forgiveness.

Superior Court Judge Todd Lang said he "loved that AI" and called it "genuine". Prosecutors had asked for nine years and the defence for seven; the judge imposed the presumptive term of 10.5 years and acknowledged it was more than prosecutors had requested.

What did the appeals court decide?

A unanimous three-judge panel, in an opinion by Presiding Judge D. Steven Williams, affirmed the conviction, vacated the manslaughter sentence and sent the case back for resentencing. Because Horcasitas had not objected to the video at the hearing, the court reviewed only for "fundamental error" — a higher bar, which it found was met.

The State argued the video was reliable because Wales knew her brother closely and the views expressed matched his own writings and beliefs; the panel was not persuaded. Comparing the video with earlier Arizona Supreme Court authority on photographs of a homicide victim's young sons at his gravesite, it drew the decisive distinction. The photographs captured a real moment. The AI video, in the court's words, "does not reflect actual events": rather than documenting an event, it presented "a depiction of the victim and his thoughts created from the imaginings of the victim's sister."

Two further findings carry the weight. First, the court held that the video "erases the interpretive distance" between what the family believed Pelkey would have said and his own voice, and that "no disclaimer could cure that error." Second, it found the video "clearly impacted" the judge; the record did not show how much, but "there is no question it played a role." The real footage embedded in the same video was permissible, the court said.

How have the parties responded?

Wales said she was disappointed the video was cited as the reason, and that her only goal had been to humanise her brother. Lawyers for the victims and for Horcasitas declined to comment, and the Maricopa County Attorney's Office said its prosecutors knew the family would address the court but were not aware of the nature of the presentation. As of 2 October 2026, no resentencing date had been set.

Judge Scott Schlegel, who writes on courts and technology, noted that the opinion does not settle how every court will treat AI recreations, and that another panel may see the issue differently.

Why the line between a record and a generation matters

This is an Arizona sentencing decision. It does not bind any court in the UK, Europe or the Middle East, and it says nothing about how those courts would approach similar material. Its value for readers there is the reasoning, which is unusually clean.

The test the court applied was not "was AI involved?" Real footage and synthetic footage sat in the same file, and the court allowed one and rejected the other. The test was whether the material documents something that happened. On appeal, Horcasitas's lawyer pressed a related point: the video did not disclose who wrote the words the AI figure spoke.

That distinction maps directly onto disputes about what AI systems say about real companies and people. An AI system's answer about a company is a generated statement. It cannot prove its own truth, and treating it as a record of fact is exactly the error the Arizona court identified. But the fact that a particular system produced that answer, to a particular question, on a particular date, is an event — and an event can be documented. That is the side of the line where evidence lives. It is also what the publishers in the Munich ruling against Google's AI Overviews relied on: a dated record of what the system said, before and after Google was told. And it is the record that, as we noted in our piece on chat logs as evidence, nobody keeps by default for the party being described.

The disclaimer finding deserves attention too. The video announced that it was AI-made, and the judge still described its words as coming from Pelkey's heart. Anyone relying on an "AI-generated" label to blunt a synthetic statement attributed to a client or an executive should not assume the label will do the work.

What this means for lawyers and comms teams

  • Separate two questions. Is what the AI system said true? And did it say it? The first usually cannot be answered by the system. The second can be answered with a record, if one was made at the time.
  • Capture the event, not just the content. The exact question, the exact answer, which system, the date and time, and the sources it cited — preserved so a third party can verify the capture was not altered afterwards.
  • Keep real and generated material separable. The Arizona court could split authentic footage from synthetic footage. A file where the two cannot be told apart invites a court to discount all of it.
  • Expect the authorship question. Who produced the words, and how, is now a question courts ask of AI material. Have the answer before someone else asks.

Sources