An Arizona appeals court overturns a 10-and-a-half-year sentence because an AI video of the victim was played at the hearing

Deepfake of a victim played in court: an Arizona appeals court overturns a 10-and-a-half-year sentence

🔎 A dead man "spoke" at the hearing. The appeals court has just answered.

May 2025, Maricopa County courthouse, Phoenix. After nine statements from relatives, a video is played: Christopher Pelkey, killed in a road rage incident in 2021, appears in it. He looks at the defendant and tells him that he forgives him. The problem: Pelkey had been dead for four years. The video was generated by AI, from his image and his voice, by his own family.

On September 30, 2026, the Arizona Court of Appeals (Division 1) answered: the 10-and-a-half-year sentence imposed on Gabriel Horcasitas, 55, is overturned. His conviction, however, still stands. The reasoning comes down to one sentence: the judge relied on an "unreliable" video, a "fundamental error" that "crossed the line" (Reuters, AP News).

Why now? Because this is the first time — to our knowledge — that a US court has overturned a sentence because of an AI-generated video. And it comes at the worst possible time: consumer video models have become good enough to fabricate a convincing dead man in an afternoon, while criminal procedure had no rules to govern it. This case will serve as a compass for dozens of courts.

The Essentials

• The facts: Gabriel Horcasitas was convicted in 2025 of the negligent homicide of Christopher Pelkey (37), shot dead during a road rage incident in 2021. At the sentencing hearing, after 9 statements from loved ones, the family played an AI-generated video in which Pelkey "forgives" Horcasitas.
• The decision: On September 30, 2026, a panel of 3 judges from the Arizona Court of Appeals vacated the sentence of 10 and a half years. The conviction stands; only the sentence falls (official opinion, PDF).
• The reasoning: The video "presents a representation of the victim and of his thoughts created from his sister's imagination" and "erases the interpretive distance" between what the family believes and what the victim would have said. And above all: "no admonition could correct this error."
• The proof of prejudice: Judge Todd Lang had described the video as "genuine", "speaking from his heart". For the court, this demonstrates that the judge treated it as testimony.
• What's next: Horcasitas will be resentenced by the Maricopa County Superior Court (BBC). And State v. Horcasitas (1 CA-CR 25-0191) becomes a ready-made citation for any defense attorney faced with a "synthetic deposition."

Recommended tools

The Pelkey video is no technical feat: it was made with tools accessible to anyone. Here is the state of the art among video generation models capable of producing this type of content (June 2025 ranking, at the time the case took a turn).

Note on scores: the values in the "Score" column correspond to the ELO ranking of the Artificial Analysis Video Arena (human comparison of generated videos), as of June 2025. The higher the score, the more the model is preferred by evaluators.

The point is not to promote these tools. It's to note that the barrier has fallen: what once required a special effects studio can now be done with a $20/month subscription and a family photo.

This is exactly what makes the legal question urgent. Technology is no longer the filter. Only the law can be — and it's arriving nearly two years late.

What the court actually took issue with in the AI video — A direct answer

Not the message — the forgiveness — but the medium: a "depiction of the victim and his thoughts created from the imagination of the victim's sister," rather than a record of a real event.

Let's reconstruct the May 2025 hearing. Nine of Pelkey's relatives testify in succession. Then the family plays the video: Pelkey, rendered in CGI, addresses Horcasitas directly, speaks to him of forgiveness, without asking for any particular sentence. Judge Todd Lang is visibly struck: he describes the video as "genuine," "speaking from his heart." Horcasitas gets 10 and a half years in prison.

The attorneys appeal. On September 30, 2026, the three-judge panel — opinion written by Judge D. Steven Williams — rules: the conviction stands, the sentence is thrown out. Two findings stand out. The video was not "reliable." And the judge did indeed rely on it: "there is no question it played a role" (Reuters).

The sentence that's going to be quoted everywhere

"Rather than document an event, the AI video presents a depiction of the victim and his thoughts created from the imaginings of the victim's sister." (AP News).

And the coup de grâce: "No disclaimer could cure that error" (official opinion).

This is the most radical part of the ruling. The problem isn't the deception, nor the absence of a label. It's the object itself: a video that attributes to the dead man thoughts he never expressed.

Note: the French glosses that translated the already-English quotes (e.g., « aucun doute qu'elle a joué un rôle ») were omitted, since they would be redundant in English.

How a grieving family came to make a deepfake

The short answer: because no one stopped them — and nothing in the procedures of the time clearly allowed for it.

Let's rewind. 2021: Pelkey is killed in a road-rage dispute that turned fatal. The trial ends in a manslaughter conviction. Then come the sentencing hearings, that unique moment in American law when loved ones may speak before the judge sets the sentence.

The family wants to pay tribute. Pelkey's sister writes a text — what she imagines her brother would have said. Rather than reading it aloud, she turns it into a video: Pelkey's image animated, his voice recreated, his words "spoken" before the court.

At this stage, no rule defines the status of such an object. It is neither evidence (it proves nothing about the facts), nor testimony (the dead cannot swear an oath), nor a simple tribute (it is submitted to the judge at the very moment he is deciding). The video passed through this gray zone without a hitch.

That is the real lesson of the case: the error did not come from a malicious process, but from an act of mourning that encountered no safeguards. Neither the prosecution, nor the defense, nor the judge had a framework for classifying this object. The appeals court has just created one for them, after the fact.

Why the "Interpretive Distance" is the Real Red Line

Direct answer: because all victim impact statements rely on this distance—and the AI video eliminated it.

Since Payne v. Tennessee (1991), the U.S. Supreme Court has allowed loved ones to speak during sentencing. But they speak on their behalf: they describe their grief, what death has destroyed, and what they imagine their deceased loved one would have wanted to say. This inference is explicit, assumed, and the judge knows they are listening to an interpretation.

TheAI video breaks this contract. It puts words into the mouth of the deceased that they never spoke, using their voice, their face, their timbre. The court puts it best: the video "erases the interpretive distance between the family's belief about what the victim would have said and the victim's voice itself."

My opinion: this is the right line, and it is finer than a simple "no deepfakes in court." A letter read by a sibling remains a testimony. A video where the deceased "speaks" becomes a performance. And an emotional performance in high definition has no equivalent in terms of persuasion.

What is still allowed, what is no longer allowed

• Permitted: oral testimonies from loved ones, letters, writings—the historical core of impact statements.
• Grey Area: montages of actual images and videos of the deceased, which document without making them speak.
• Now Contested: any voice or visual synthesis attributing words to the deceased, even if presented as a tribute.

Important clarification: the court is not saying the family did wrong. They believed they were paying tribute, and the video did not even request specific penalties. What the court is saying is that the judge could not rely on this object to determine a sanction. The nuance is enormous for the future: the target is not the intention, it is the probative use.

Technology has made the question inevitable

Direct answer: in 2021, when Pelkey was killed, making a dead person speak convincingly was studio-level work. In 2025, it's an afternoon's work with tools costing less than $30/month.

Voice cloning requires only a few seconds of recording — and Pelkey, like everyone else, had left voice messages and videos. Animating a still photo has become a standard feature of video models. At the time the video was created, Google's Veo 3.1 had just introduced synchronized native audio — spoken dialogue generated alongside the image — while Kling 2.0 Pro and Dreamina Seedance 2.0 dominated the leaderboards.

And it goes further. Google is now pushing any-to-any models like Gemini Omni, which accept text, image, audio, and video as input to generate video. Translation: "casting" a deceased person from a family photo and three voice messages is no longer an R&D project. It's a prompt.

The reliability paradox

Here's the paradox of our era: AI has just proven a geometry theorem that had resisted for 80 years, yet a court rules that an AI video is not "reliable".

A contradiction? No. Two different kinds of reliability. A theorem can be verified by anyone, step by step. A dead man's "thoughts" can be verified by no one. That's exactly the court's argument: there is no way to verify that these words are Pelkey's, or that he would ever have said them.

Digital watermarks (SynthID from Google DeepMind, the C2PA standard) change nothing about this: they prove that content is generated, not that it is true. And the court preemptively brushed the argument aside — "no disclaimer could cure that error". Transparency doesn't repair the absence of reliability. It accompanies it, that's all.

Toward national case law?

Direct answer: yes, and probably fast — because federal deregulation is leaving courts and states to face these questions alone.

In Washington, the doctrine is "innovation first": the AI safety executive order was rescinded, and no one is waiting for a federal framework on the use of generative AI in criminal proceedings. The result: AI common law will be written decision by decision, state by state. State v. Horcasitas is the first milestone in that body of law.

Who will use it? Defense attorneys first: the citation is ready to deploy against any AI-generated victim video. Prosecutors next, who will avoid this kind of exhibit like the plague — losing a conviction over an emotional effect is a career risk. And finally, judges, who will have to write rules: mandatory disclosure, format limits, even outright bans.

Jurisdictions that haven't ruled yet will take a position. Most will fall in line: no one wants to be the next court to see its sentence overturned over this.

The uncomfortable angle: victims' rights

Victims' families will see it as a silencing. The court answered in advance: the loved ones' voice remains intact — letters, testimony, public tributes. What is forbidden is lending the voice of the dead. It's a defensible distinction, but it will fuel a genuine political debate in the states, where victims' lobbies are powerful.

A quick historical note for perspective: American courts had already stumbled over generative AI. In 2023, a New York lawyer cited case law invented by a chatbot (the Mata v. Avianca case). But this is the first time a sentence has been vacated because of AI content submitted directly to the judge. The scale of the harm has changed.

And what it doesn't settle: the next front will be elsewhere. A defense producing a generated "alibi"? An expert relying on a model to establish a probability of guilt? The underlying problem — the evidentiary reliability of an AI artifact — remains fully intact. Arizona has just drawn the first line.

❌ Common Mistakes

Mistake 1: Believing a disclaimer is enough

Slap an "AI-generated video" overlay on it and everything's fine? The Arizona court settled the matter: "no disclaimer could cure that error." The problem isn't missing information — it's the effect produced. The solution: treat the medium as a carrier of meaning, not just the content. A warning doesn't cancel out an emotion.

Mistake 2: Confusing tribute with testimony

Posting an AI video of a departed loved one on social media, as a tribute, is one thing. Submitting it to a judge is another. The solution: strictly separate the contexts. What belongs to private remembrance must never claim the status of evidence submitted to a court's decision.

Mistake 3: Lending someone's voice or face without a framework

Voice cloning of a real person without consent is already being challenged in several states — Tennessee criminalized it as early as 2024 with the ELVIS Act. The solution: written consent, or abstention. For the deceased, consent is by definition impossible. That's precisely why the line is red.

Mistake 4: Thinking that "the judge knows it's fake, so it doesn't influence them"

That's the Pelkey family's implicit reasoning — and the court invalidated it. The voice of the dead creates an authority that nothing can counter. The solution, on the court's side: rely not on the judge's discernment, but on rules of admissibility. On the citizen's side: treat any talking video of a real person as potentially synthetic.

❓ Frequently Asked Questions

Is Horcasitas's conviction overturned?

No. The court of appeals upholds the conviction for negligent homicide. Only the 10.5-year sentence is vacated, because the judge relied on the AI video. Horcasitas remains guilty in the eyes of the law, but he will be resentenced at a new sentencing hearing.

What exactly did the video say?

Christopher Pelkey appeared in it as a computer-generated likeness, addressing Horcasitas directly and telling him that he forgave him. The video did not ask for any specific sentence. It was played in May 2025, after the statements of nine relatives, and had been created by the family using the victim's image and voice.

On what legal basis did the court rely?

On the concept of "fundamental error": an error so serious that it justifies vacating the sentence. The court finds the video "unreliable" — it documents nothing, it "presents the victim's thoughts created from his sister's imagination" — and notes that the judge used it to set the sentence.

Does this ban AI videos in all courts?

No. This is a decision from an Arizona court of appeals: it binds the courts of that state and influences others without compelling them. But that is precisely what makes it a precedent. Every American jurisdiction will now have to take a position on this line — and most will fall in line.

Can relatives still use AI to speak on behalf of a deceased loved one?

In the courtroom, no — that is the heart of the decision. Outside of court, nothing changes legally in most states, except where deepfake laws apply. The court's position is clear: writing or testifying on behalf of the deceased, yes. Making him speak with his voice and his face before a judge, no.

When will Horcasitas be resentenced?

No specific date has been announced. The case goes back to the Maricopa County Superior Court, where the judge will have to impose a new sentence — without the video, or within a framework strictly defined by the rules stemming from the appellate decision.

✅ Conclusion

An appellate court has just set the first real legal limit on generative AI in a courtroom: you can speak for the dead, but you cannot make them speak. This case also shows how these questions are now being decided outside of labs and courts—in what everyone can produce, publish, and share from their browser. If you are closely following these changes and want to publish your own analyses as decisions are made, a simple hosting solution like Hostinger is enough to launch your site (offers starting from a few euros per month, check hostinger.com).