Opinion: An AI ghost spoke at sentencing — and a court was right to silence it

On Sept. 30, the Arizona Court of Appeals vacated the sentence of a man convicted of killing a stranger in a 2021 road-rage shooting — not because the jury got the verdict wrong, but because the judge who sentenced him had been shown a video of the dead man forgiving him. The video was not real footage. It was an artificial-intelligence re-creation, scripted by the victim's sister and voiced by a synthetic likeness of a man who could no longer speak for himself. The appeals court called it unreliable and "fundamentally unfair." It was right to say so, and it should not have taken this long for a court anywhere in the country to say it.

A Video Built From Grief, Not Fact

Christopher Pelkey was shot and killed at a Chandler, Arizona, intersection in 2021 after a confrontation with Gabriel Paul Horcasitas, who was later convicted of manslaughter after a jury rejected his self-defense claim. At Horcasitas's May 2025 sentencing hearing, Pelkey's sister, Stacey Wales, played an AI-generated video in which an avatar built from her brother's photos and voice told the courtroom he believed in forgiveness and that, "in another life," he and his killer might have been friends. Wales has said she wrote the script herself, without AI drafting tools, after struggling to find words for a traditional victim-impact statement.

It worked — on the judge, at least. Maricopa County Superior Court Judge Todd Lang called the video "genuine" and said he was moved by what he described as the victim's "obvious forgiveness." He then sentenced Horcasitas to 10.5 years — roughly a year more than the nine-and-a-half years prosecutors had asked for, in part, the appellate record shows, because of what the synthetic Pelkey had said.

The Appeals Court Drew a Necessary Line

In State v. Horcasitas, a three-judge panel of Arizona's Court of Appeals, Division One, left the manslaughter conviction untouched but ordered a new sentencing hearing, finding that the AI video was not a record of anything that had actually happened and should never have carried the evidentiary weight it was given.

"The AI video so prejudiced Horcasitas as to render the sentencing procedure fundamentally unfair," the court wrote, adding that the technology "erases the interpretive distance between the family's belief of what the victim would have said and the victim's own voice and opinions."

That distinction matters. Arizona's victims' rights laws — rooted in the state constitution and codified at A.R.S. § 13-4426.01 — give crime victims a genuine and hard-won right to be heard at sentencing, free from cross-examination. But that right belongs to the people in the room: survivors, family, the living. It was never meant to license a stand-in, however lovingly constructed, to testify to feelings the dead person never actually expressed. A grieving family's sincere belief about what their brother would have wanted is not evidence of what he actually believed, and a judge is not supposed to sentence a defendant based on a eulogy dressed up as testimony.

The Rulebook Still Hasn't Caught Up

What should trouble readers beyond Arizona is that this case is not an aberration the system is equipped to handle — it is a preview of a gap nobody has closed. The same federal judiciary that would ordinarily be expected to set national guardrails for exactly this kind of evidence has been unable to finish the job. A proposed Federal Rule of Evidence 707, meant to hold machine-generated evidence to the same reliability standard as expert testimony, cleared publication for public comment in mid-2025 — but after the comment period ended, the Standing Committee on Rules of Practice and Procedure declined to advance it in June 2026, sending the proposal back for further study alongside a separate, unresolved fight over deepfake evidence.

That leaves a patchwork in which:

• Arizona's own ruling is a published appellate opinion binding within that district, not a statewide rule written for every courtroom
• No federal evidence rule yet governs AI-generated evidence of any kind, including at sentencing, where rules of evidence are already relaxed
• Judges are left to assess, case by case and often for the first time, whether a synthetic video, voice clone or avatar is "reliable enough" to sway a sentence
• Victims' families have every incentive to keep producing this kind of material, because — absent a rule saying otherwise — it plainly works

Judge Lang is not a cautionary villain here; he is a preview of how any sympathetic judge might react when grief is rendered persuasively enough. That is precisely the problem with leaving this to instinct rather than rule.

What the Horcasitas Case Should Force Courts to Do

None of this requires banning families from using technology to express loss, and the appeals court did not do that — genuine photos and recordings of Pelkey remain fair material for a victim-impact statement. The argument for a hard rule is narrower and more defensible: synthetic recreations that put invented words in a dead person's mouth should never be treated, consciously or not, as if they were the victim speaking. Courts that allow such material at all should require it to be clearly labeled as an advocacy tool authored by survivors, kept separate from any portion of the record a judge is told reflects the victim's own statements, and explicitly excluded from the kind of individualized weight Judge Lang gave it. That is a modest, achievable standard — and it is exactly the kind of guardrail that federal rulemakers spent a year drafting and then shelved rather than finished.

The Horcasitas case will likely be cited for years as the first time an American court confronted an AI "witness" after the fact. The more useful legacy would be if it were also the last time a court had to figure this out after the sentence was already handed down. Arizona's appellate judges did the hard part — naming the problem precisely and refusing to let a conviction's punishment rest on invented testimony. The unfinished work belongs to rulemaking bodies, state and federal, that have had this exact scenario in front of them for more than a year and still have not written the rule that would have prevented it.