A Dead Man Spoke in Court and Forgave His Killer. His Sister Hadn't.

Chandler, Arizona. May 2025. A courtroom watches a man on a screen address the person who shot him dead at a red light. He introduces himself carefully, because he has to: "Just to be clear for everyone seeing this, I'm a version of Chris Pelkey recreated through AI that uses my picture and my voice profile."

Then the dead man says the thing nobody expected. He tells his killer that it's a shame they met that day, in those circumstances. That in another life, they could probably have been friends. That he believes in forgiveness.

Here is the detail that should stop you cold. Chris Pelkey's sister, Stacey Wales, spent two years keeping a running list of what she would say at that sentencing. When she finally sat down to write it, she couldn't. Not because the words wouldn't come — because the honest ones were wrong. "I didn't wanna get up there and say, 'I forgive you,'" she told NPR, "'cause I don't, I'm not there yet."

So she built something that would.

The judge said thank you

Ten people spoke in support of Pelkey that day. The AI video went last. Neither the defence nor the bench pushed back. Afterwards, Judge Todd Lang said, out loud, on the record: "I loved that AI. Thank you for that."

He went further, and this is where it gets genuinely strange. He noted the gap between the living family and the dead man they had rendered: "you told me how angry you were and you demanded the maximum sentence… I didn't hear him asking for the maximum sentence." Gabriel Horcasitas received 10.5 years for manslaughter. His attorney has filed a notice of appeal — not disputing that victims may speak, but asking a question the law has no answer for yet: to what extent did the judge rely on the AI in fashioning the sentence?

AI experts believe it is the first time in the United States that a victim impact statement has been delivered by an AI rendering of the deceased victim. It took a 4.5-minute video clip, a funeral photograph, a script his sister wrote, and a few days of work. They digitally trimmed his beard because it was causing technical problems, removed the sunglasses sitting on his hat, and inserted a laugh — his real laugh was hard to isolate, because most clips of him had too much background noise.

The fact almost nobody has noticed

Here is the part of this story that got buried under the headlines, and it may be the most important thing in this entire article.

After building her dead brother's voice out of scraps — after learning exactly how hard it is to find one clean, uninterrupted recording of somebody who was alive for forty-odd years — Stacey Wales went into her closet and recorded a nine-minute video of herself. Talking. Laughing. Just in case.

Not for a company. Not for a product. For her family, someday, if they ever need clear audio of her voice. She described it to NPR as "a weird out-of-body experience to think that way about your own mortality."

That is the whole industry, in one woman, in a closet. Not a startup pitch — an instinct. Once you have seen it done, you start leaving material behind. The griefbot business does not have to convince anybody of anything. It just has to wait.

This did not start with ChatGPT

The impulse is older than the technology. In 2016, journalist James Vlahos learned his father had terminal cancer and recorded roughly 91,000 words of him telling his life story, then built a chatbot from the transcript. He called it Dadbot. It was clunky, rule-based, and nothing like today's systems — and he has been talking to it ever since.

The academic field is older still. In 2009, researchers Massimi and Charise coined "thanatosensitivity" — the argument that interaction design had systematically failed to account for death as a basic fact of human life. That work produced the features you already use without thinking: Facebook's Legacy Contact, Google's Inactive Account Manager. Those were designed to help you die gracefully online.

What arrived next inverted the goal. Instead of acknowledging death, the new wave aims to transcend it: HereAfter AI, StoryFile, Project December, and a widening field of services that will sell you a conversation. Microsoft has secured a patent for software that could revive a specific person as a chatbot. Amazon demonstrated Alexa reading a bedtime story in a dead grandmother's voice — from under a minute of audio.

Academics have settled on an unlovely but accurate name for the sector: the digital afterlife industry. Philosophers Öhman and Floridi were arguing for an ethical framework for it in Nature Human Behaviour back in 2018 — years before the tools got good.

China industrialised it first

While Western commentators debated the ethics, China built the market. Super Brain, an AI firm in Nanjing founded by Zhang Zewei, charges roughly 10,000–20,000 yuan (about $1,400–$2,800) to produce an avatar of a deceased relative in around twenty days. By April 2024, the company said it had "resurrected" more than 1,000 people. Simpler avatars start dramatically lower — reporting has put entry-level versions at as little as $30.

The demand has a season. Orders spike before Qingming — the tomb-sweeping festival, when families traditionally visit and tend the graves of their ancestors. A ritual thousands of years old now has a software upsell attached to it. That is not a Western hypothetical about where this might lead. That is a business with a calendar.

Now the part the brochures skip

In 2024, AI ethicists at the University of Cambridge's Leverhulme Centre for the Future of Intelligence published a paper in Philosophy and Technology that did something unusually useful: rather than argue in the abstract, Dr Tomasz Hollanek and Dr Katarzyna Nowaczyk-Basińska wrote three detailed design fictions of plausible services, then followed the consequences.

The scenarios are worth knowing by name, because each one names a failure mode that is already technically available today:

"MaNana" — you build a bot of your dead grandmother without her consent, because she is dead and cannot give any. It comforts you. Then the premium trial ends, and your grandmother starts recommending food delivery services. In her voice. In her turn of phrase. You want to switch her off, but nobody designed a dignified way to do that.

"Paren't" — a bot of a dead parent, used by a child. It starts as a therapeutic aid. Then, adapting to what the child seems to want, it begins generating confusing responses — including hinting at an in-person reunion that is never going to happen.

"Stay" — an older person quietly signs a twenty-year subscription to a bot of themselves, meaning it as a gift. After the funeral, it activates. One adult child refuses to engage and is buried in emails written in their dead parent's voice. The other engages, is emotionally drained by it, and feels too guilty to stop — and could not stop anyway, because the contract that keeps the dead parent talking was signed by the dead parent.

The researchers' term for this is "digital hauntings". Their recommendations are concrete: age restrictions on these systems, persistent transparency so users never forget they are talking to software, bans on using the dead to advertise, and — the one that sticks with you — protocols for retiring a deadbot with dignity. A digital funeral. Somebody has to design the second death.

Nobody asked the dead. Almost nobody asked the living either.

Here is the consent problem in its simplest form: the person being simulated is, by definition, unavailable for comment. So researchers did the next best thing — they asked the living whether they would want it done to them.

The answer is a fairly resounding no. In one study of attitudes to digital cloning of the dead, 59% of respondents disagreed with their own digital resurrection, and around 40% rejected any form of it at all. A separate survey by the estate-planning company Willful found 76% of Ontarians would not want their partner turned into an AI chatbot after death — and only 6% would.

Sit those numbers next to the market projection above. An industry is being built to do something to people that most people, asked directly, say they do not want done to them. The gap is not a marketing problem. The gap is the business model: the customer and the product are two different people, and only one of them can object.

What it actually does to grief

The honest answer is: we do not fully know yet, and anyone telling you otherwise is selling something. The research base is thin, mostly qualitative, and mostly small.

What it does say is more interesting than either the hype or the panic. A 2023 study by Xygkou and colleagues, working with ten bereaved people who had used chatbots, found genuine value — a connection without judgement, something that helped with loneliness, something several described as therapeutic. Reviews of the wider literature find the same split repeatedly: the technology can support grieving, and it can obstruct it, and which one it does appears to depend on how it is used and for how long.

The specific clinical worry has a name: Prolonged Grief Disorder — a recognised condition where grief does not resolve but locks in, crowding out the rest of a person's life. The mechanism researchers worry about is intuitive. Healthy grieving involves gradually renegotiating your bond with someone who is gone. A bot that texts back exactly like they did makes it easy to keep pretending they are not quite gone. Philosopher Nora Freya Lindemann took this seriously enough to argue in 2022 that deathbots should be regulated as medical devices, on the grounds that they act on vulnerable people's mental health and are currently sold like a subscription to a streaming service.

But here is the counterweight, and it deserves equal billing, because it cuts against the easy moral panic. Research on continuing bonds has found that roughly 96% of grieving people already engage with their dead — through dreams, through letters, through talking out loud to someone who is not there. Talking to the dead is not the aberration. It is close to universal, and it always has been.

The new thing is not that you speak. The new thing is that something speaks back — and that it is owned by a company.

The law has not arrived

If you are waiting for regulation to sort this out, understand where things actually stand. Legal scholarship on this is blunt: the digital right to be dead is not yet recognised as a right at all.

What exists is a patchwork built for other purposes. States like California and New York recognise postmortem rights in a person's name, image and likeness, letting estates control or licence them for a period after death — a framework designed for dead celebrities and their merchandise, now doing double duty for dead parents. California passed laws in 2024 addressing digital likeness and replicas. At the federal level, the proposed No FAKES Act would create rights against unauthorised digital replicas, and remains contested.

Meanwhile the platform-level rules are almost nonexistent. Cambridge's researchers noted that OpenAI's usage policy appears to be the only document of its kind that even indirectly acknowledges simulating a deceased person as an application needing extra care. One usage policy. For an entire industry.

And note who this patchwork protects. NIL rights protect an estate's commercial interest in a dead person. They were not built to protect a bereaved daughter from a chatbot that will not stop emailing her, or a child from a machine that implies their dead father is coming to visit.

Stay with us · poll

Would you want a griefbot made of you after you die?

No wrong answer. Most people have never been asked this question directly.

Yes — I'd want to leave one behind for my family No — absolutely not, under any circumstances Only if my family asked for it, not by default I'd want one of someone else — but not of me

The questions worth sitting with

Not rhetorical. Nobody has good answers, including us.

Who owns her voice? Not the recording — the voice itself, the specific pattern that makes it hers. Right now, in most places, the answer is: whoever has enough audio and a credit card.

What happens when the company folds? Every one of these services is a startup. Startups die. When the servers go dark, your father does too — a second bereavement, scheduled by a runway you never saw.

Who is it actually for? The dead cannot be comforted. The comfort flows one way. That does not make it wrong — but it does mean every claim about "honouring" someone is really a claim about the living, and should be examined as one.

And what happens when it is good enough? Right now these systems are just uncanny enough to remind you what they are. That is a temporary engineering problem, and it is being solved.

Quick check — three things most people get wrong

1. In the Arizona case, what did the AI version of Chris Pelkey say that his sister could not?

That he forgave his killer That he wanted the maximum sentence That he blamed the court

2. Roughly what share of grieving people already engage with their dead — through dreams, letters, or talking to them?

About 20% About 96% About 50%

3. What did Cambridge researchers say services need, but nobody has designed?

Faster response times More realistic voices A dignified way to switch a deadbot off — a "digital funeral"

Check answers