The Challenge
A trial court in Arizona admitted an AI-generated video of a murder victim delivering his own impact statement at sentencing. The victim's sister created the avatar using artificial intelligence to simulate her brother's voice, facial expressions, and thoughts about forgiveness. The sentencing judge praised the presentation and imposed the maximum presumptive sentence of 10.5 years, exceeding both the prosecution's request for 9 years and the defense's request for 7 years.
The Arizona Court of Appeals vacated the sentence and ordered a new hearing, ruling that the AI video was unreliable and fundamentally unfair. This case highlights how generative AI can blur the line between permissible evidence and prejudicial fabrication.
The Environment and Constraints
The trial court operated under Arizona's victim impact statement framework, which allows family members to describe how a crime affected them and the victim. Courts routinely admit photographs, letters, and video recordings that capture actual moments or documented events.
Previously, the Arizona Supreme Court addressed admissibility limits in a case involving photographs of a victim's young sons at his gravesite. The court found the images permissible because they documented a real moment experienced by surviving victims, despite being "uncomfortably close" to undue prejudice.
In this case, the trial judge was informed about how the AI video was created. The video included embedded footage of the actual victim alongside AI-generated segments. Despite these disclosures, the presentation suggested the victim himself was delivering the statement.
The Approach Taken
The trial court admitted the AI video as victim impact evidence. The judge acknowledged the family's anger while highlighting the AI avatar's message of forgiveness, then imposed a sentence longer than the prosecution recommended.
The defendant challenged the admission on appeal. The appellate court applied a fundamental error analysis, comparing the AI video to the previously approved gravesite photographs.
The Court of Appeals made a clear distinction: "Unlike the photographs in Rose, which captured a particularly poignant moment for the young surviving victims, the AI video here (excluding the embedded 'real' videos of the victim) does not reflect actual events. Indeed, 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."
Results and Metrics
The appellate court vacated the sentence and remanded for resentencing. The court identified three specific failures:
Authentication failure. The AI video presented fabricated content as the victim's own voice and opinions. The avatar stated, "this here is a true representation of who I was, you know, not how the courtroom portrayed me." No actual basis existed for attributing these specific words or sentiments to the deceased.
Interpretive erasure. The technology eliminated the gap between what family members believed the victim would say and what he actually would have said. Traditional victim impact statements from family members preserve that distance; the court hears the family's perspective, not a simulation of the victim's mind.
Demonstrated prejudice. The appellate court noted the AI video "clearly impacted the sentencing judge," who imposed a sentence exceeding both parties' recommendations after praising the AI presentation.
The embedded authentic video footage of the victim remained permissible. The court's objection targeted only the AI-generated segments that fabricated the victim's statements and expressions.
What They Would Do Differently
The appellate opinion doesn't explicitly address what the trial court should have done, but the analysis points to three procedural controls:
Apply existing authentication standards. Courts already require proponents to establish that evidence is what it purports to be. An AI avatar isn't the victim. Treating it as such fails basic authentication requirements regardless of how clearly you label it.
Distinguish documentation from fabrication. Real videos of the victim, photographs, and family members' own statements about their loss all document actual events or perspectives. Generative AI creates new content that never existed.
Assess prejudicial impact separately from disclosure. The trial court knew the video was AI-generated. That disclosure didn't cure the prejudice because the presentation format still attributed fabricated statements to the victim. A disclaimer can't fix structural unreliability.
Takeaways for Your Team
If you're advising on evidence production or litigation readiness, this case exposes three gaps in current practice:
Your authentication protocols don't address generative AI. Most legal hold and preservation policies assume you're collecting documents and communications that already exist. When a party or witness creates new content using AI, you need protocols that distinguish between: (1) authentic records created by the custodian, (2) AI-assisted drafting where the human controlled the output, and (3) AI-generated content that fabricates statements or data points.
Metadata alone won't establish reliability. The trial court had full disclosure about how the AI video was created. Technical metadata showing "this file was generated by [AI tool]" doesn't answer whether the content is admissible. You need substantive review of what the AI fabricated versus what it documented.
Generative AI creates recordness problems you can't solve with disclaimers. A record represents evidence that something happened or someone said something. When AI generates the content, you're not preserving a business record; you're creating a simulation. Your Records Control Schedule probably doesn't have a retention class for "AI-fabricated statements attributed to deceased persons," and it shouldn't.
The Arizona appellate court didn't ban AI from courtrooms. It applied existing evidence rules to new technology and found that fabricated content failed reliability standards. That's the framework you should apply when evaluating whether AI-generated materials belong in your litigation hold scope or production set.
If you're collecting it, you should be able to authenticate it. If you can't authenticate it without inventing what someone would have said, you're not managing records; you're creating fiction.
