Court Vacates Pelkey Sentence After Judge Said He Liked AI Video of the Victim

A court has vacated the sentence of Gabriel Horcasitas, convicted of killing Christopher Pelkey in Arizona. In May 2025, Judge Todd Lang sentenced him to 10.5 years, more than the 9.5 years the prosecution requested. Before that, the courtroom screened about four minutes of AI video in which the victim himself "spoke." The judge afterward said he liked it. That remark is now part of a story that ended with the sentence being vacated.

An important caveat. From the original link (an NBC News report), we have only the headline, and we do not have the court's full reasoning. What follows are the verified circumstances of the 2025 case and our own analysis, not a summary of the court's ruling.

What happened in the courtroom

Pelkey, a 37-year-old US Army veteran, was shot in 2021 in Chandler after a road-rage confrontation. Horcasitas was found guilty of manslaughter. At sentencing, the victim's family has a legal right to speak about the harm done to them. Pelkey's sister, Stacey Wales, went beyond the usual letters and speeches: together with her husband and an acquaintance who works in the field, she built an AI avatar of her brother from photos and voice recordings.

Wales wrote the script herself. In the video, the avatar said that in another life he and the defendant might have been friends, and expressed a belief in forgiveness. The judge called it sincere and then imposed the maximum sentence.

Why this is a legal problem

The problem has three layers.

• Authorship of the words. The dead man said nothing. A living person wrote the lines, yet the viewer hears the victim's voice and sees his face. Such a clip carries more emotional weight than a statement read aloud.
• No cross-examination. The defense cannot test whether the avatar's "position" matches Pelkey's actual views. Victim statements cannot be challenged the way witness testimony can, and here a synthetic "voice" took their place.
• The judge's reaction. A remark that he liked the clip sits poorly with the requirement of impartiality. For an appeal, it is a ready-made quote in the record.

We do not know how exactly the court justified vacating the sentence. But the defense's logic is clear: the sentence rests on material that looks like testimony but is not, and the judge publicly showed that it moved him.

Case metrics

• Sentence imposed in 2025: 10.5 years.
• Prosecution's request: 9.5 years.
• Length of the AI video: about 4 minutes.
• Status: the sentence is vacated, and the case returns to the question of punishment.
• Year of the crime: 2021.

What courts, lawyers, and developers should do

For courts, the takeaway is simple: they need written rules on synthetic material at the sentencing stage. At a minimum, those rules should require disclosing AI use in advance, showing the clip to both sides before the hearing, labeling the video and its script, and giving the defense a chance to object.

Lawyers should decide in advance how to respond to AI inserts: a motion for a preview, a substantive objection, and putting the court's reaction on the record. Remarks like the judge's are what later carry weight on appeal.

Product teams building voice- and video-cloning tools should start thinking about "grief and remembrance" scenarios. Ethics warnings, watermarks, and a content provenance log can help show that the material is synthetic and who wrote the text. Otherwise, such cases will be settled without developers at the table, and probably not in their favor.