Reliable, Not Genuine

On Wednesday the Arizona Court of Appeals, Division One, vacated a manslaughter sentence because the sentencing judge had relied on an AI-generated video of the victim. The ABA Journal’s report of it ran under the headline “Judge wrongly permitted AI-generated victim video,” and some of the coverage I saw reads as if a court had drawn a line against AI in the courtroom. I read the 16-page published opinion, State v. Horcasitas, No. 1 CA-CR 25-0191, filed 30 September 2026. The court holds something narrower, and I think more durable.

It did not rule on notice, or oaths, or video. It ruled on reliability, and the word the opinion turns on is not “AI” but “reliable,” and the word it quotes the judge using is “genuine.”

What happened in the room

Gabriel Horcasitas shot and killed Christopher Pelkey during a road-rage confrontation in November 2021. A jury rejected his self-defense claim and convicted him of manslaughter. At sentencing, which the AP covered in May 2025, Pelkey’s family played a video in which an AI rendering of Pelkey, built from his picture and a profile of his voice, addressed the court and the defendant. The opinion describes it in detail. The avatar opens by saying it is a “version” of the victim “recreated through AI.” It later says the video “is a true representation of who [he] was.” It tells Horcasitas, “I believe in forgiveness and in God who forgives.” Real footage of Pelkey is embedded in it twice.

The words were not Pelkey’s. Just before the video played, his sister told the court what she was about to do: “what would he have said if he could have taken my place today?” The court describes the video as “created from the imaginings of the victim’s sister.”

Then the sentencing judge responded to it. He said, “I loved that AI. Thank you for that,” and, “I feel that that was genuine,” and, “I didn’t hear him asking for the maximum sentence.” He then imposed 10.5 years, which, because the State had alleged no aggravating factors, was the presumptive term and the most he could give (¶35). The prosecutor had asked for 9. The judge noted that this was “more than the State asked for.”

What each party asked for at the Horcasitas sentencing, and what the court imposed, in years A dot plot of manslaughter sentence length in years, on an axis from 6 to 11. The defense asked for 7 years. The State asked for 9 years. The victim's family, through counsel, asked for the maximum authorized by the verdict, which was 10.5 years. The court imposed 10.5 years. The axis starts at 6, not zero. MANSLAUGHTER SENTENCE, YEARS · STATE V. HORCASITAS, ¶¶35–41, 30 SEP 2026 Defense asked State asked Victim's family, via counsel Court imposed 7 years — defense request, ¶40 9 years — State request, ¶38 10.5 years — the maximum authorized by the verdict, requested by family counsel, ¶¶35, 37 10.5 years — imposed, ¶¶13, 41 7 9 10.5 10.5 6 7 8 9 10 11 Axis starts at 6 years. The family's request was "the maximum sentence authorized by the jury verdict," which the opinion puts at 10.5.
The avatar spoke of forgiveness, and the sentence landed at the ceiling anyway. The opinion does not say which way the video pushed the number.

Reading what the court did and didn’t say

The figure is the first thing to notice, and I would not over-read it. A video of a forgiving victim, followed by the maximum sentence, looks like the video failed to move the judge. The court does not say that. It says “the record does not reflect precisely how the AI video factored into the sentencing calculus,” and also that “there is no question it played a role” (¶52). The judge’s remarks tie the video to his view of the family (“it also says something about the family”) and to his sense of the victim’s character, and I can’t tell from the opinion’s excerpts what he would have done without it. The court did not try to quantify the video’s effect either. It found that relying on it, on this record, so prejudiced the defendant that the sentencing was fundamentally unfair (¶53).

That is why the holding is about the evidence rather than the outcome. Sentencing judges have wide latitude in what they hear; the opinion cites Arizona’s rule that “[a]ny reliable, relevant evidence” is admissible, but also the limit that “only evidence bearing a ‘minimal indicia of reliability’” may be considered, and that defendants have “a due process right to a fair sentencing procedure which includes the right to be sentenced on the basis of accurate information” (¶45). Against that, the State argued the video was reliable because it reflected what the sister believed her brother would have said, because the siblings were close, and because it matched his written statements (¶47). The court’s answer is in ¶51, and I’ll quote its core, since it is the paragraph the rest turns on:

“despite the explanations concerning the genesis of the AI video, it presents the statement (including non-verbal communications made through facial expressions and the like) as coming directly from the victim. 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.”

The sister’s belief about what he would have said is a legitimate thing for a victim’s family to tell a court. The opinion says nothing against it. The trouble is the rendering. It puts a belief in a face and a voice, and the sentencing judge treated it as the man’s own forgiveness. The court adds, flatly, “No disclaimer could cure that error.” The video opened with a disclaimer. The judge still called it genuine.

What the Horcasitas opinion rested on, and what it did not Two panels. The left lists what the court relied on: the AI video does not reflect actual events (paragraph 50); it presents the statement as coming directly from the victim (paragraph 51); no disclaimer could cure the error (paragraph 51); the sentencing judge called it genuine (paragraphs 41 and 52). The right lists what the opinion did not turn on: advance disclosure of victim statements is not required (paragraph 42); oaths and cross-examination are not required of victims (paragraph 42); the real footage embedded in the video was permissible (paragraph 53). A note says review was only for fundamental error because the defense did not object (paragraph 34). STATE V. HORCASITAS · PARAGRAPH REFERENCES TO THE OPINION WHAT IT RESTED ON WHAT IT DID NOT TURN ON The video "does not reflect actual events" (¶50) It presents the words as the victim's own (¶51) "No disclaimer could cure that error" (¶51) The judge called it "genuine" (¶41, ¶52) No advance disclosure: not required (¶42) No oath, no cross-examination: not required (¶42) The real footage inside the video: permissible (¶53) No objection was made at sentencing (¶34). Review was for fundamental error only; the court found it.
The holding is a reliability holding. The procedural points people reach for first are, in the opinion, background or left untouched.

The procedural points the court left alone

The ruling does not rest on any of the procedural points people might reach for first. Victims have a constitutional right to be heard at sentencing, and the opinion notes that they “are not required to disclose their statements to the State, the defendant, or the court,” do not take an oath, and are not cross-examined (¶42). It also says the real footage of Pelkey embedded in the video “was permissible” (¶53). So the court does not touch the unsworn, undisclosed, or recorded character of victim statements. It carves out the synthesized part and nothing else.

One wrinkle I’d flag, and it is my own reading, not the court’s. Because victims need not disclose statements in advance, the first time defense counsel sees an AI rendering may be in the courtroom. The defense here did not object at sentencing, so the court reviewed only for fundamental error (¶34). It found that error. But the route to it is a narrow one, and a defendant whose lawyer is surprised by an avatar in the room is in a hard place to object in time. I have no data on how often that happens, and the opinion does not discuss it.

The court was working without precedent. “No Arizona case has addressed the admissibility of an AI-generated depiction of a victim offered as victim impact evidence” (¶48). The authority it reaches for is a 2013 state supreme court case on two photographs of a victim’s sons at his grave, State v. Rose, which the justices called “troubling” and “uncomfortably close” to the line before upholding. The distinction the panel draws is that the photographs recorded something that had happened. The AI video, excluding its embedded footage, “does not reflect actual events.” I looked for any official 2026 guidance on AI in victim impact statements and did not find any in the sources I checked. For now, a published opinion from one Arizona panel is most of the law there is, and the court itself calls the issue “novel.”

A defendant tried twice, and now to be resentenced, over a dead man’s words

One detail from the opinion’s background explains how this defendant came to stand before that judge. He was tried twice. After a first trial, the superior court found the State had withheld text messages from the victim’s phone, in violation of Brady and the state’s disclosure rule, and threw out the guilty verdicts, entered acquittals on the murder and drive-by charges, and ordered a new trial. The Court of Appeals affirmed that order in an earlier memorandum decision, which the new one cites (¶¶6–8). The second jury convicted of manslaughter. On this appeal, the defense also argued the second trial wrongly excluded those texts, and the panel rejected that argument and affirmed the conviction (¶1).

I mention it because so much of this case turned on what a court could know about what a dead man thought. At the first trial the dispute was over his texts. At sentencing it was over a video of him speaking. The questions differ, but in both someone had to decide how far a record of the man could stand in for him.

That is also why I’d read this case alongside the Penske and Chegg dismissal from earlier this week: in both, the press headline compressed a court’s narrow reasoning into a broad verdict on AI. Here the narrowness is the point. The panel did not say AI has no place near a victim’s family. It said a sentencing judge cannot treat a synthetic voice as the victim’s own, and that putting “this is AI” at the top of the video does not fix that. Whether judges, prosecutors and defense lawyers elsewhere will adopt that line, and whether Arizona’s supreme court will take the case, I don’t know. A resentencing is now ordered; its outcome is not in any source I could find.

References

  1. State v. Horcasitas, No. 1 CA-CR 25-0191 (Ariz. Ct. App., Div. One). Opinion, Williams, J., filed 30 September 2026. All paragraph citations and quotations in this piece are from this opinion. No reporter citation was available on the filed PDF.
  2. ABA Journal (2026). Judge wrongly permitted AI-generated victim video, state appeals court says. 1 October 2026. Source for the November 2021 date of the shooting (attributed there to Reuters) and the 10.5-year term.
  3. Matt York, Rio Yamat and Sejal Govindarao (2025). AI-generated video gave victim a voice at his killer’s sentencing in Arizona. Associated Press, via News4Jax, 8 May 2025. Background on the original hearing.
  4. State v. Rose, 231 Ariz. 500 (2013). Cited in the opinion (¶¶34, 44, 48–50) as the governing victim-impact precedent.