The dead cannot testify. AI should not speak for them

Christopher Pelkey had been dead for more than three years when his face and voice appeared before an Arizona court.

He was shot during a road-rage confrontation in November 2021.

At the sentencing of Gabriel Horcasitas in May 2025, Pelkey's sister presented an artificial intelligence (AI)-generated recreation of her brother.

The figure addressed Horcasitas directly, spoke about faith and forgiveness, and claimed to be a true representation of the man Pelkey had been.

The video openly acknowledged that it was artificial. The judge knew this, yet described the apparent forgiveness as genuine.

On September 30, 2026, the Arizona Court of Appeals upheld the manslaughter conviction but vacated the sentence in State v Horcasitas.

It found that the sentencing procedure had been fundamentally unfair.

The sentencing court had given a family's interpretation of the victim's wishes the authority of the victim himself.

Although the judgment arose under American law, the problem extends well beyond its courts.

Public authorities regularly consider the wishes of people who cannot speak for themselves, particularly in care and safeguarding decisions.

Relatives may provide valuable evidence of someone's preferences, but a synthetic portrayal could make their interpretation more persuasive without making it more reliable.

Public institutions should not have to experience the same failure before addressing that risk.

A close relationship does not establish reliability

The State defended the video's reliability by pointing to the family's close relationship with Pelkey and their knowledge of his character.

Pelkey's brother had also read from something Pelkey wrote during his lifetime. Those were his actual words, unlike the new words attributed to him in the AI video.

The family was entitled to explain what they believed Pelkey would have wanted.

The sentencing judge nevertheless treated the AI-generated portrayal as a personal expression from Pelkey, although Pelkey had never expressed forgiveness towards Horcasitas.

The appellate court found that the reconstruction erased the "interpretive distance" between the family's understanding of Pelkey and his own voice and opinions.

It presented newly created words and expressions through his likeness.

In State v Rose (2013), the Arizona Supreme Court considered photographs of a murdered police officer's young sons at his gravesite.

The Horcasitas judges distinguished those images, which depicted events that had occurred, from a reconstruction of a deceased person's thoughts.

The sentencing judge told the family he "loved that AI" and regarded the apparent forgiveness as genuine. The appellate court found that the video influenced his decision, although it could not establish precisely how.

Its finding on the video was blunt: "No disclaimer could cure that error."

Is labelling alone enough?

Labelling can tell an audience that content was artificially generated. It cannot establish whether the words or wishes attributed to somebody are genuinely theirs.

Article 50 of the EU AI Act, applicable since 2 August 2026, requires disclosure of certain deepfakes and other AI-generated content.

The European Commission's guidance excludes personal, non-professional use from the Act's deployer obligations, meaning a family acting privately could fall outside those requirements.

The proposed US NO FAKES Act, reported by the Senate Judiciary Committee in June 2026, addresses rights over replicated voices and likenesses, including after death. It has not become law.

These measures address disclosure and control of digital identity.

Neither determines whether a court or public agency should accept invented words as evidence of someone's wishes. Consent to reproduce a voice cannot establish consent to the words placed in it.

Arizona already had a legal standard capable of addressing the problem.

Defendants have a right to due process and to be sentenced on reliable information. The difficulty was that existing procedures did not require advance scrutiny of the synthetic material before it influenced sentencing.

Victims have constitutional rights to be heard, construed liberally.

Under Arizona law, their statements are not ordinarily subject to advance disclosure to the prosecution, defence or court. Nor are victims required to take an oath or submit to cross-examination.

The defence had not objected to the video. However, it had no general entitlement to receive the victim-impact material beforehand. The appellate court nevertheless found fundamental error.

The sentencing judge imposed ten and a half years, the presumptive term and the maximum available without alleged aggravating factors, rather than the nine years requested by the prosecution.

The appellate court found that the video so prejudiced Horcasitas that sentencing was fundamentally unfair, although it could not determine precisely how the portrayal affected the sentence.

Existing law provided a remedy, but only after the sentence had been imposed. It had supplied the standard without a reliable opportunity to apply it beforehand.

Courts should not receive synthetic first-person statements as an individual's own words or wishes unless that person actually communicated the specific statement, or expressed the specific view, about the matter in question.

General beliefs or an apparently consistent character are insufficient.

Had that rule applied in Horcasitas, Pelkey's genuine writings and footage could still have been considered.

His family could still have described their loss and their understanding of his beliefs. The invented expression of personal forgiveness could not have been attributed to him.

Making that restriction effective would require advance notification when a victim-impact presentation includes a synthetic representation of an identifiable person.

Courts would then have an opportunity to establish who authored the words and whether the material is being presented as the individual's own views.

Victims' rights to be heard are constitutionally protected, and Arizona does not ordinarily require their statements to be disclosed beforehand.

A limited notification requirement for synthetic material would introduce an additional obligation and may require changes to procedural rules.

The right to be heard has limits

The right to be heard should not be confused with an unrestricted right to determine what material a court must accept.

Court rules committees and judicial authorities could address synthetic victim-impact presentations through procedural rules and practice directions.

Ministries of justice and other public agencies could introduce equivalent guidance when receiving material used in decisions about people's rights or welfare.

The restriction should not prevent someone using synthetic speech to communicate words they genuinely authored or authorised.

That person is preserving their own voice.

A relative describing what an absent person might have wanted is offering their interpretation and should be heard on that basis.

The sentencing court already had Pelkey's words, genuine footage and his family's account of their loss.

The synthetic reconstruction added the appearance of the victim personally responding to the man who killed him.

Pelkey never gave that response. The dead cannot correct the record. That makes the responsibility not to invent their words greater, not smaller.

The author writes in a personal capacity and his writing does not reflect the views of his employer or GovInsider.

------------------------------------------

The author is a UK public sector practitioner with more than 17 years’ experience across policing, investigations, safeguarding, operational leadership and risk. He holds a First-Class BSc (Hons) in Applied Criminal Justice and is studying an MA in International Relations, Security and Strategy. He writes on AI governance, organisational learning and the integrity of public-sector decision-making.