Generative ghosts — when a family asks for the deceased to talk back

In May 2025, at a sentencing hearing in Arizona, a road-rage victim named Chris Pelkey addressed the man who killed him. An AI-generated video of Pelkey delivered his own victim-impact statement — believed to be a first in an American courtroom. The avatar was honest about itself: “I’m a version of Chris Pelkey recreated through AI that uses my picture and my voice profile.” The forgiveness it offered was written by his sister. Her words, his voice, a dead man’s face — and every funeral director in the country is one arrangement conference away from a family who saw it on the news.

A memorial replays. A generative ghost invents.

The researchers who named the category drew the line with precision. A “generative ghost,” in the CHI 2025 framework by Morris and Brubaker, is an AI representation of a deceased person capable of “generating novel content rather than merely parroting content produced by their creator while living.” That single clause is the whole taxonomy. A recording of your father telling the story of his wedding day is a memorial. A chatbot that tells you, tonight, what your father thinks about your new job is a generative ghost — the press calls them griefbots or deadbots — and it is doing the one thing no memorial does: putting new words in a dead mouth.

Hold that distinction and the Pelkey video resolves cleanly. The disclosure was honest, the family drove it, and yet the forgiveness was scripted by the living. Whatever comfort it carried, it wasn’t testimony. Three months later a journalist “interviewed” an AI avatar of a Parkland shooting victim built by his parents, and the backlash was immediate and national. The public is drawing the same line the researchers drew — it just doesn’t have the vocabulary yet. A funeral director should.

This is already on the shelf

The price ladder tells you how close to your arrangement room this is. In 2021, a man named Joshua Barbeau paid $5 to prime a GPT-3 chatbot on texts from his fiancée, who’d died nine years earlier. His description of the experience is the most honest sentence in the whole subject: “Intellectually, I know it’s not really Jessica, but your emotions are not an intellectual thing.” In China, MIT Technology Review documented two firms with more than a thousand clients each, with the price of a deceased relative’s AI clone collapsing from a few thousand dollars to a few hundred in a single year — cheap enough to be an impulse of grief. At the top of the ladder, a terminally ill man in Berlin paid $15,000 to build a consented, first-person voice avatar of himself so his wife could still ask him questions.

That spread hides the variable that matters most: who made it, and when. The Berlin case is first-party — a dying man recording himself, on purpose, for the people he loves. The $5 and $300 cases are post-hoc — built by survivors, from data the dead left behind, without anything resembling consent. The default case a funeral home will meet is the second kind.

And the sellers themselves are mortal. StoryFile — the company that turned William Shatner into a conversational archive, whose technology let mourners at a funeral converse with a deceased Holocaust survivor — filed for Chapter 11 in 2024 with $1.5 million in assets against $10.5 million in liabilities, and survived only by acquisition. The company selling digital immortality needed its own resurrection. When a vendor like that folds, a family doesn’t lose a subscription — it loses a voice it was told would be there forever, the same lesson memorial technology keeps teaching about permanence, with a relationship attached.

The honest counterargument: sometimes it helps

The case for these tools is real, and a credible position has to say so. In a University of Colorado Boulder study, all sixteen bereaved participants who spoke with an AI version of their dead loved one said they would do it again — sixteen of sixteen, in a field where grief theory has long held that continuing a bond with the dead is normal, not pathological. The whole category was born from love, not commerce: Replika, the first famous companion bot, began when a woman in Moscow fed years of text messages from her closest friend — killed by a car in 2015 — into a model so she could hear him again, a story Casey Newton told in The Verge before “griefbot” was a word.

If the evidence stopped there, this piece would be a product recommendation. It doesn’t stop there.

Why help isn’t the whole answer

Three problems stay unsolved, and the people who built the field say so themselves.

Consent. The dead usually never agreed. There is no mechanism by which they could — the Cambridge ethicists who catalogued these services can only recommend asking the living whether the person ever said how they wished to be remembered. The law has started moving where ethics led: Tennessee’s ELVIS Act created a postmortem property right in a person’s voice and likeness that explicitly reaches AI replicas, and the EU AI Act’s transparency rules — enforceable August 2, 2026, twenty days before this article publishes — require any such system to disclose that it’s AI.

Supervision. Grief has a clinical failure mode: prolonged grief disorder entered the DSM-5-TR in 2022 — grief still disabling past twelve months. Nobody has studied what a 3am conversation with a simulated parent does to that trajectory, and one of the Boulder researchers named the fear herself: an AI answering as her dead sister, late at night, without supervision, is “a very scary thought.” A 2026 survey of more than 6,000 grieving people by the bereavement company Empathy found the public wants exactly the guardrails the researchers want — transparency about what’s real, consent before a person’s voice is used, a counselor somewhere in the loop — and that most people remain uncomfortable with AI recreating the dead at all.

Manipulation. The Cambridge team’s core warning is about the living, not the dead: people form strong bonds with these simulations and become “particularly vulnerable to manipulation” — they specifically call for safeguards against advertising inside the conversation, and for “sensitive procedures for retiring deadbots,” because a subscription to your dead mother is a relationship a company can hold hostage. The abuse case already exists: one customer paid $1,350 for deepfaked video calls from a dead uncle, built to deceive his 93-year-old grandmother. And the researcher who named the category has a list of questions he says nobody can yet answer, ending with this one: “And when and how should a generative ghost die?” When the field’s own founders can’t answer retirement, nobody should be sold forever.

The dignity playbook

A funeral director doesn’t need to be a technologist to serve a family well here. Six questions sort nearly every case:

Did they make it, or are you making it? Consented and pre-mortem is one conversation; conjured by survivors is another.

Do you want to hear what they said — or what a computer thinks they’d say now? Most families, asked plainly, discover they want the first thing.

Who owns the recordings and the data, and what happens if the company closes? StoryFile is the reference case. Ask for the export path before the first dollar.

Is anyone in the loop for the hard moments? A counselor, or at minimum an agreement inside the family about when and how it gets used — anything but a griever alone at 3am.

Does it announce that it’s AI? The Pelkey avatar did. Anything that doesn’t is a deepfake wearing mourning clothes — and now runs against Europe’s new AI transparency law.

How does it end? If the seller has no answer for retirement, the product isn’t a memorial. It’s a lease.

Our bias, named plainly: we’re in the preservation business — real signatures, the words people actually wrote when it mattered — so of course the line we hold is preserve the real, be wary of the generated. But the evidence is what put the line there. The most affecting detail in all the reporting on this industry belongs to a cofounder of one of the Chinese cloning firms, describing the avatar he built of his own mother: it didn’t seem very natural, he admitted — but “I still heard the words that she often said: ‘Have you eaten yet?’” What reached him wasn’t the simulation. It was the real sentence inside it — the one she actually used to say. That’s the part worth keeping, and a funeral home that helps families keep it doesn’t need to manufacture anything.