Arizona court throws out a sentence over an AI victim video

The Arizona Court of Appeals has thrown out a manslaughter sentence after finding that a trial judge should never have allowed an AI-generated video of the victim to be played at the sentencing hearing. The three-judge panel vacated the 10.5-year prison term on September 30 and sent the case back to the Maricopa County Superior Court to fix the sentence. The conviction itself stands.

The hearing it undid took place in May 2025. Nine of the victim's family members and friends had already stood up to describe what the killing had done to them. Then a second video was played: a synthetic version of the man who was killed, speaking in his own voice, forgiving the man who shot him, and telling him that "in another life, we probably could have been friends". The victim was a 37-year-old Army veteran, killed in November 2021 during a road-rage confrontation at a stoplight in Chandler, about 25 miles outside Phoenix. He was unarmed and had got out of his truck to walk toward the other driver's vehicle, according to the Associated Press.

The victim's sister wrote the script from what she believed he would have said; she and her husband, with a family friend, built his likeness from a photograph run through an "old age" filter. It did not ask for any particular sentence. It is believed to be the first AI-generated victim impact statement played in a United States courtroom.

After it played, the trial judge said he "loved that AI" and called the presentation "genuine", adding that it gave the victim a chance "to speak from his heart". He then imposed the prison term. In the opinion released Wednesday, Judge D. Steven Williams wrote for the panel, joined by Judge Anni Hill Foster and Vice Chief Judge David Weinzweig: "While the record does not reflect precisely how the AI video factored into the sentencing calculus, there is no question it played a role."

The court was blunt about why that mattered. "Rather than documenting an event or recording a particular moment, the AI video presents a depiction of the victim and his thoughts created from the imaginings of the victim's sister," the panel wrote. It held that the judge's "consideration of the AI video so prejudiced [the defendant] as to render the sentencing procedure fundamentally unfair".

Prosecutors had defended the clip, telling the appeals court that it accurately represented the victim's character and that a close family relationship allowed "the reasonable inference that the sentiments conveyed by the AI victim were probably accurate". The defence argued the video never disclosed who wrote the words in it, was backed by nothing that established the victim actually held those views, and carried an "undue emotional weight" that a family member reading the same lines aloud would not have had.

The court also noted how little settled law there was to work with: "no Arizona case has addressed the admissibility of an AI-generated depiction of a victim offered as victim impact evidence." It said it was guided by an earlier Arizona Supreme Court decision that had found photographs showing a murdered officer's sons at his grave came "uncomfortably close" to being unfairly prejudicial. The panel drew the line there: those photographs recorded a real moment, and the AI video did not.

In a statement, the Maricopa County Attorney's Office said its prosecutors knew the victim's family would address the court but were not aware of the nature of what would be played.

Arizona is not the first court to push back on synthetic media. In 2024 a Washington state judge excluded an AI-upscaled bystander video from a murder trial, holding that the tool had not been accepted by the forensic video analysis community, which is the standard that state applies to novel scientific evidence, according to a Greenberg Traurig analysis. At the federal level, the evidence rules committee has proposed a new Rule 707 that would apply expert-evidence reliability gatekeeping to machine-generated evidence; as of this year's Chambers practice guide it has not been adopted, and the committee is separately weighing an amendment to Rule 901 that would make the authenticity standard stricter for evidence that may have been fabricated with AI, though it is still surveying courts on how prominent that problem is.

The practical lesson lands on anyone who makes synthetic depictions of real people for an official purpose. A court will ask who authored the words, what evidence establishes they reflect the person's actual views, and whether the format gives them an authority the words alone would not carry. A recording of the dead speaking in their own voice, in a room where the dead cannot be cross-examined, is unlikely to clear that bar. A family member reading the same statement aloud is.

The resentencing returns to Maricopa County Superior Court. A victims' lawyer has told the family that using AI again could produce another appeal; the sister has said she will keep trying to have her brother's voice heard "in whatever allowable medium" the court permits, the AP reported.

Resentencing has not been scheduled; the appeals court left the conviction in place, so the case returns only for the sentence.