Should a Killer's Victim Be Able to "Speak" at a Sentencing Through AI?

An Arizona trial court judge allowed this innovative approach to presenting a victim impact statement, which seems like a useful step toward justice.

Last week, the road-rage killer of an Arizona man was sentenced to 10 1/2 years in prison after his victim "spoke" to the court via a victim impact statement (VIS) created through artificial intelligence. The trial court judge allowed the victim's family to play an AI-generated video with a version of the victim—including his face and body and a lifelike voice—which appeared to ask the judge for leniency. The family has posted the statement here on YouTube. The defendant was apparently sentenced to the maximum term and has promised to challenge the statement on appeal. In my view, this kind of statement seems like a useful technological innovation that should be generally allowed, subject to reasonable limitations imposed by the trial court (which I discuss at the end of this post).

In this post, I discuss this AI-generated victim-impact statement against the backdrop of  the existing legal landscape in this country. All fifty states and the federal system allow victims to deliver victim impact statements, as I outline in this article. (Interestingly, many other countries use some form of victim impact statement as well.) In homicide cases such as this Arizona case, the victim's family members step into his shoes as his representative to deliver the statement. I have made the general case for allowing victim impact statements in two papers, "In Defense of Victim Impact Statements" and "How Victim Impact Statements Promote Justice: Evidence from the Content of Statements Delivered in Larry Nassar's Sentencing" (co-authored with Professor Edna Erez). The U.S. Supreme Court has also approved victim impact statements even in capital cases, in Payne v. Tennessee (1991). To be sure, academics such as Professor Susan Bandes and Professor Mike Vitiello have raised objections to VISs. But for purposes of this post, I will assume that the legitimacy of VISs  and address the narrower question of whether an AI-generated statement should be permitted.

This question should be assessed in light of the recognized purposes of victim impact statements. The first purpose is to provide relevant information to the sentencers, often (as in this Arizona case) a judge. This purpose has been described as the "informational rationale" for victim impact statements. Through a VIS, the victim's family members are allowed to provide "a quick glimpse of the life" the defendant "chose to extinguish," thereby reminding the sentencer that "the person whose life was taken was a unique human being," as Justice O'Connor explained in her concurring opinion in Payne. An AI-generated statement simply builds on that approach.

An AI-generated statement is not too far removed from previously used technologies, which help to provide the glimpse into the victim's life. For example, in homicide trials, it has long been the practice to allow the prosecution to introduce a photograph of the victim taken when the victim was alive. For example, a Utah statute (enacted in 1994 to enforce Utah's Victims' Rights Amendment) provides that "[i]n any homicide prosecution, the prosecution may introduce a photograph of the victim taken before the homicide to establish that the victim was a human  being, the identity of the victim, and for other relevant purposes." Of course, in a criminal case by the time of sentencing, the defendant has been proven guilty beyond a reasonable doubt, and so issues related to potential bias while determining guilt are no longer in play.

At sentencing, appellate courts have approved of video montages of photographs of a victim and even videos of a victim. For example, in 2009, the California Supreme Court considered the admissibility in a death penalty case of an eight-minute video of the victim and his family enjoying a trip to Disneyland. The Court rejected a defendant's challenge, calling the tape an "awkwardly shot 'home movie'" lacking the elements generally designed to stir up emotions." People v. Dykes, 209 P3d 1, 44-45 (Cal. 2009).

However, such videos can go too far. For example, in 2004, a federal district court excluded a victim video in a death penalty case. The court explained that the video, made for a memorial  service, "was about twenty-seven minutes in length and featured over 200 still photographs of the victim, in roughly chronological order, from the time he was born until the time just before his death. The pictures were set to evocative contemporary music, including that of the Beatles and James Taylor." This video, the court concluded, when too far "because its probative value was outweighed by the danger of unfair prejudice, and created a danger of provoking undue sympathy and a verdict based on passion as opposed to reason." United States v. Sampson, 335 F. Supp. 2d 166, 191 (D. Mass. 2004). At the same time, however, the district court noted that other courts had allowed some video clips to be used, but in "each of these cases, the admitted video was brief and found to be probative of some aspect of the victim's life."

An AI-generated statement is simply a technological advance on this tradition of victim videos, and thus helps to address one of the core problems that VIS are designed to address. If no victim statement is allowed, it turns the killer's victim "into a faceless stranger" and thus "deprives the State of the full moral force of its evidence," as the U.S. Supreme Court recognized in Payne. Through the AI-generated statement in this Arizona case, the victim—Christopher Pellkey—was no longer a "faceless stranger" to the proceedings but rather a unique human being whose life the defendant extinguished.

In determining whether an AI-generated statement should be allowed, it is also useful to compare two situations. Suppose the defendant has seriously injured a victim by striking a non-deadly blow. In that case, the victim himself is allowed to deliver a victim impact statement about what happened. For comparison, suppose that the defendant has killed a victim by intentionally striking a lethal blow. In that case, the victim is obviously unavailable to deliver his own statement. An AI-generated statement is simply a technological substitute for the fact that the defendant has criminally chosen to make the victim unavailable. In balancing the equities between an innocent victim and a killer, of course, the equities lie with the victim and his family representatives. Because the defendant has chosen to commit his deadly crime, any doubts should be resolved against him in deciding whether an artificially created substitute for the victim is an appropriate response to the victim's absence .

This fairness point was emphasized by Justice Souter's in his concurring opinion in Payne, where he explained:

Murder has foreseeable consequences. When it happens, it is always to distinct individuals, and, after it happens, other victims are left behind. Every defendant knows, if endowed with the mental competence for criminal responsibility, that the life he will take by his homicidal behavior is that of a unique person, like himself, and that the person to be killed probably has close associates, "survivors," who will suffer harms and deprivations from the victim's death.
Just as defendants know that they are not faceless human ciphers, they know that their victims are not valueless fungibles; and just as defendants appreciate the web of relationships and dependencies in which they live, they know that their victims are not human islands, but individuals with parents or children, spouses or friends or dependents. Thus, when a defendant chooses to kill, or to raise the risk of a victim's death, this choice necessarily relates to a whole human being and threatens an association of others, who may be distinctly hurt. The fact that the defendant may not know the details of a victim's life and characteristics, or the exact identities and needs of those who may survive, should not in any way obscure the further facts that death is always to a "unique" individual, and harm to some group of survivors is a consequence of a successful homicidal act so foreseeable as to be virtually inevitable.

A second purpose for victim impact statements is to serve expressive and communicative functions that can produce therapeutic benefits for victims (and, in homicide cases, for their families). This therapeutic rationale for victim statements is well supported by the field of therapeutic jurisprudence or "TJ." And the AI-generated statement in the Arizona case illustrates this purpose well. In the statement, "Christopher Pellkey" (as reconstructed through AI) gets to give his last goodbyes to his family and friends, in the formal setting of the sentencing of his killer. Pellkey's AI-generated last words are haunting, but might reasonably be expected to provide some small measure of comfort to his family and friends:

Thank you to everyone for being here. It means more than you know. Well, I'm gonna go fishin' now. Love you all. See you on the other side.

A third purpose for victim impact statements is to explain the full harm of the defendant's crime—to the defendant. This purpose is unrelated to the sentence that is ultimately imposed in the case, but rather rests on the consequences of a victim looking the "defendant in the eye and let[ting] him know the suffering his misconduct has caused" (as the Ninth Circuit as explained in its Kenna decision, 435 F.3d 1011, 1016 (9th Cir. 2006)). To be sure, eye-to-eye contact between victim and defendant is not physically possible in a homicide case. But an AI-generated substitute might begin to serve the same purpose of encouraging the defendant to work toward rehabilitation.

This particular Arizona victim impact statement provides a good illustration. In the statement, the reconstructed Pellkey speaks directly to the defendant:

To Gabriel Horcasitas, the man who shot me: It is a shame that we encountered each other on that day in those circumstances. In another life, we probably could have been friends. I believe in forgiveness and God who forgives. I always have. And I still do.

The final purpose of a victim impact statement is serving a public educative function. Even critics of such statements have conceded that they can potentially work towards this goal. For example, Professor Bandes admits that VISs might serve to "call[] attention to crimes that are poorly understood and underenforced." Here again, Pellkey's AI-generated statement illustrates this point. The statement powerfully illustrates the far-reaching harms and lasting consequences of road rage escalations. Hopefully someone, somewhere will view that statement on YouTube and think twice before escalating to deadly force. Of course, that hope is speculative, and no one knows for certain whether the statement will produce any public benefit. But it seems worth a try.

When an AI-generated victim impact statement is allowed in a homicide case, the trial court should scrutinize it to ensure it meets several criteria:

1. The statement should be reasonably related to describing the impact of the crime, as that is an underlying purpose of the statement. In other words, the statement should be focused on providing "a quick glimpse of the life" the defendant chose to extinguish.
2. The statement should not personally attack the defendant or speak to issues beyond the scope of a legitimate victim impact statement. For example, the Utah Supreme Court reversed a sentence in murder case where the victim's family members used their victim impact statement to call the defendant a "terrorist" who could not be rehabilitated. The Court concluded that these characterizations were unduly prejudicial and went beyond the bounds of an appropriate victim impact statement. State v. Ott, 2010 UT 1. An AI-generated statement would need to respect such boundaries as well.
3. The statement should focus on the victim who was killed without inviting comparative judgments to others in the community. One of the criticisms of VISs is that they invite the sentencer to decide who is a "worthy" and "unworthy" victim. Payne responded to this concern, explaining that "victim impact evidence is not offered to encourage comparative judgments of this kind—for instance, that the killer of a hardworking, devoted parent deserves the death penalty, but that the murderer of a reprobate does not. It is designed to show instead each victim's uniqueness as an individual human being." An AI-generated statement would need to follow that path as well.
4. The statement should respect any restrictions imposed in the jurisdiction or the  particular case about whether a sentencing recommendation by the victim or victim's family is allowed. In capital cases, for example, current Supreme Court precedent (Bosse v. Oklahoma) does not permit a victim's family to give an opinion, one way or the other, about whether a death penalty should be imposed. An AI-generated statement should not be used to circumvent that limitation.
5. Finally, the statement should follow whatever reasonable time limitations and other restrictions the trial judge might choose to impose. The Arizona statement was less than four minutes long. In my research on the Nassar victim impact statements, Professor Erez and I found that the sex assault victims' statements averaged about 1,200 words or eight minutes in length. Of course, what is a reasonable length would need to be assessed by the trial judge handing a particular case against the backdrop of that case and the judge's docket management needs.

One additional concern an AI-generated statement raises is whether the technology will be available to all victims' families, regardless of their financial circumstances. It is unclear how much this Arizona statement cost to produce. As AI-technology becomes more broadly available, no doubt the cost to create such a statement will fall substantially. But in the immediate short term, prosecutor's offices and victim service organizations should be alert to resource disparities and try to arrange pro bono or other low-cost help to victim's families who wish to use such statement.

As it advances, AI will offer important innovations in many areas of law and life. The desirability of each of those innovations will need to be assessed on its own merits. The innovation of an appropriate, non-prejudicial AI-generated victim impact statement appears to be a positive one.