On the first day of May in 2025, a dead man stood up in a courtroom in Maricopa County, Arizona, and forgave the person who killed him. Christopher Pelkey had been shot at a red light near Gilbert and Germann roads in 2021, in the kind of stupid, irreversible road-rage encounter that ends a life in seconds. Four years later, at the sentencing of the man convicted of killing him, the courtroom watched a video of Pelkey looking out from the screen and speaking in something close to his own voice. “To Gabriel Horcasitas, the man who shot me, it is a shame we encountered each other that day in those circumstances,” the figure said. “In another life, we probably could have been friends. I believe in forgiveness, and a God who forgives. I always have, and I still do.”
Pelkey said none of this, of course. He was dead before the sentence existed. The words were written by his sister, Stacey Wales, who had spent two years drafting what she wanted to tell the court and found that the only voice she could hear clearly was her brother's. She and her husband trained generative AI on old photographs and a single video clip, reconstructed his face and his voice, and let the simulation deliver a message of mercy that the real man never got the chance to refuse. The presiding judge, Todd Lang, told the room he loved it. Legal scholars, watching from a distance, felt something closer to vertigo.
The vertigo is the point. A commercial and increasingly accessible technology can now reconstruct a person after death, animate their likeness, approximate their manner, and put words in their mouth, and the entire apparatus operates in a zone where almost nothing is settled. Who consented to this. Who controls it. Who profits. What happens when the simulation says something the dead person would have found repugnant. These are not edge cases waiting for a future crisis. They are live questions, being answered ad hoc, case by case, by grieving families and the companies that sell to them, while courts and legislatures stand at the edge of the problem and squint.
In April 2026, three researchers tried to map the squinting. Their paper, published in the journal Philosophy & Technology, is titled “The Many Faces of Indeterminacy in Interactive Deadbots,” and its central claim is unnervingly precise. The technology that simulates the dead does not merely raise hard questions. It sits inside a structural fog, an indeterminacy so deep and so multi-dimensional that the usual instinct, to wait for the law to catch up, may be a category error. There might be nothing, in the current frameworks, to catch up.
The Industry That Sells the Voices of the Dead
Start with what is actually for sale, because the commerce is the part most people still find hard to believe.
A “deadbot,” in the now-standard if grimly cheerful vocabulary of the field, is an AI system that simulates a deceased individual using their voice, their likeness, and the digital traces they left behind. The terms multiply like anxieties: griefbots, thanabots, ghostbots, postmortem avatars. They sit inside what Cambridge researchers have named the digital afterlife industry, and that industry is no longer a thought experiment. Estimates of its scale vary by methodology, but Zion Market Research valued the broader digital legacy market at roughly 22.46 billion US dollars in 2024, with other analysts projecting growth into the tens of billions across the coming decade. Whatever the exact figure, the direction is unambiguous. Mourning has become a market.
The products differ in ambition. At the more restrained end sits HereAfter AI, a US company that records a living person through guided interview sessions and turns those recordings into an interactive “Life Story Avatar” that family can later question. The person doing the recording chooses what to include. The result is closer to an interactive memoir than a séance, an archive that answers back. StoryFile, founded in 2017 and best known for transforming the actor William Shatner into a conversational video that audiences could interrogate, took a similar interview-led approach, layering natural-language software over pre-recorded footage so that a visitor at a memorial could ask questions and receive answers assembled from the deceased's own words.
StoryFile is also a cautionary tale about the fragility of the whole enterprise. In May 2024 the company filed for Chapter 11 bankruptcy protection in the Southern District of New York, declaring around 1.5 million dollars in assets against some 10.5 million dollars in liabilities. It later emerged from bankruptcy after its assets were acquired by a new owner. Sit with that sequence for a moment. The repository of a dead person's reconstructed self, the thing a family paid for so they could keep talking to their mother, becomes a line item in a creditors' schedule, an asset to be sold to whoever wins the auction. The continuity of the dead, in this model, depends on the solvency of a start-up.
At the more aggressive end of the market are systems built to generate rather than replay. Project December, constructed on early OpenAI models, lets users summon a chatbot of more or less anyone by feeding it text samples and a personality sketch. You, Only Virtual asks for the raw sediment of a specific relationship, the text threads and voice notes, and produces a “Versona” you can message and call. Seance AI works from described traits and writing styles. The distinction matters enormously. A replayed archive can only say what the person said. A generative model says new things, in the dead person's voice, that the dead person never said and might have hated. The Philosophy & Technology paper calls this technological indeterminacy, and it argues, crucially, that it is not a bug to be patched. Large language models are nondeterministic by design. Bias, hallucination and opacity are not teething problems on the way to a faithful resurrection. They are structural features of the medium. The dead, reconstructed this way, will always be capable of saying something untrue to who they were.
Five Kinds of Not Knowing
The paper's three authors, Atay Kozlovski and Roel Dobbe of TU Delft in the Netherlands, and Edina Harbinja of Birmingham Law School, have between them a useful combination of expertise. Harbinja in particular has spent years building the legal scholarship on what she calls post-mortem privacy, the question of whether the dead retain any protectable interest in the data they leave behind. Their argument is not the familiar one that deadbots are creepy, or that grief should be sacred, or that Silicon Valley has gone too far. It is colder and more structural than that. They identify five distinct dimensions along which interactive deadbots are indeterminate, and they show how the dimensions feed one another.
The first is technological, the inherent unpredictability of generative systems already described. The second is social. Grief, the authors note, has no single correct shape. It varies across individuals and cultures, across faiths and families, and the same interface that consoles one mourner may corrode another. By industrialising grief through what they call algorithmic mediation, deadbots impose a uniform commercial product on a deeply non-uniform human experience, and there is no settled standard for telling healthy use from harmful use. The third dimension is philosophical, and it is the one that quietly destabilises everything else. What, metaphysically, is the relationship between the simulation and the person it imitates. Is it a representation, a continuation, a puppet, a corpse made of words. Can a user ever know whether the thing is telling them something the dead person believed, or merely something statistically likely given the training data. These are not rhetorical flourishes. They determine whether harm is even possible, and to whom.
It is the fourth and fifth dimensions, the legal and the regulatory, where the abstraction becomes urgent and where the original question sharpens to a point. Because here the indeterminacy is not philosophical hand-wringing. It is a measurable absence of law.
The Law That Stops at the Graveside
European data protection is often held up as the strongest privacy regime on the planet, the framework that forces global companies to bend. It is also, on the specific matter of the dead, almost entirely silent.
Recital 27 of the General Data Protection Regulation states the position with brutal economy. The GDPR “does not apply to the personal data of deceased persons.” The reasoning runs deep into the structure of the right. Data protection in the European tradition is a personal right, attached to the living individual, and on the standard view it is extinguished at death along with the person. The rights that operationalise it, the right to be informed, the right of access, the right to erasure, the right to object, all of them require a data subject to exercise them, and a data subject is, by definition, alive. When your mother dies, her data does not inherit her protections. It simply stops being protected.
This is the gap that the deadbot industry occupies, and it is not an accident that the products exist there. The same recital that closes the door leaves it slightly ajar, providing that member states “may provide for rules regarding the processing of personal data of deceased persons.” A handful have walked through. France is the clearest case. Article 85 of its data protection law lets any person issue directives, before death, about the retention, deletion and communication of their personal data afterwards, and where no instructions exist, the heirs step into the role. France has gone further still. In late 2025 its data protection authority, the CNIL, devoted its tenth Innovation and Foresight Report, titled “Our Data After Us,” to precisely this terrain, examining everything from account transmission to the new commercial offering of deadbots, conversational agents trained on the deceased, and calling for clearer rights and regulation of AI applied to post-mortem data. The French National Digital Ethics Council has urged specific supervision of systems “purposely imitating the way of speaking or writing of a deceased person.”
The United Kingdom, by contrast, offers almost nothing. There is no general statutory post-mortem privacy right. What governs the fate of your digital remains is, overwhelmingly, the contract you clicked through without reading, the terms of service of whichever platform holds your data, interpreted through a patchwork of property, intellectual property, succession and probate law that was never designed for the question. Research led by Harbinja and colleagues, surveying more than 1,700 UK adults, found a strong public appetite for control over digital remains coexisting with almost no awareness of, or use of, the few tools that exist. People want to decide what happens to them after death. They do not know that, legally, they mostly cannot.
The United States is fragmented in its own way. Post-mortem publicity rights, the right to control the commercial use of a person's identity, survive death in some states, notably California and New York, but they were built for celebrities, for estates with a brand to monetise. They protect the commercial value of a dead person's identity rather than the dignity of an ordinary one. A famous musician's estate can sue over an unauthorised hologram. The family of a private individual whose voice has been cloned into a chatbot has, in most jurisdictions, no equivalent claim, because the law sees no market value to defend, and dignity, in this corner of the legal system, has never quite counted as an injury.
Is the Simulation a Person, a Thing, or Neither
Underneath the patchwork lies a problem the paper names with real precision. Post-mortem law occupies unstable ground between persons and things, and an interactive deadbot refuses to settle on either side.
Consider what a deadbot simultaneously is. It is a creative work, a piece of software and recorded media that someone authored and might own under intellectual property law. It is a dataset, an assembly of personal information that data protection regimes might, in principle, govern, except that the regimes stop at death. And it is an extension of a personality, a representation of a specific human self that touches on dignity, reputation and privacy. Each of those categories pulls toward a different legal owner and a different body of rules. The work belongs to its author, perhaps the company. The data belonged to a person who no longer legally exists. The personality belonged to the dead, whose interests the law struggles to recognise once they are gone.
So the question of who controls the thing has no clean answer, and the paper shows how that control fragments in practice. It scatters across platforms and providers, families and communities, none of whom hold complete authority. Families have what the authors call affective stakes, and in some jurisdictions limited legal ones, but the platforms function as what they memorably describe as de facto co-authors of the past. A policy shift, an API change, an algorithmic update, a bankruptcy, any of these can erase an archive, distort its provenance, or quietly rewrite the narrative of who someone was. The dead do not get a vote. Often the living barely do.
This is why the consent question is so much harder than it first appears, and why “anticipatory” frameworks like consent-by-proxy or stewardship, which governance discussions increasingly invoke, do not dissolve it. The deceased's actual preferences, whether they wished to be revived at all and if so how and by whom, are, in the paper's words, often simply unknown. Pre-mortem consent, where the person records themselves while alive, as with HereAfter AI, gets you closest to something defensible, but even there the consent is necessarily incomplete. You can agree to be remembered. You cannot meaningfully agree, in advance, to every new sentence a generative model will one day produce in your name, because neither you nor anyone else can know what those sentences will be. Consent to a process whose outputs are structurally unpredictable is a strange and attenuated kind of consent. It is closer to a leap of faith than a contract.
When the Dead Speak in Public
The deepest discomfort arrives when the reconstructed dead are deployed not for private solace but for public argument, because there the gap between what the person said and what the simulation says becomes a matter of contested record.
In August 2025, the former CNN correspondent Jim Acosta published an interview with an AI-generated avatar of Joaquin Oliver, who was murdered at the age of seventeen in the 2018 Parkland school shooting. The avatar was created by Oliver's parents, who have spent years campaigning for gun reform, and it appeared on what would have been their son's twenty-fifth birthday. On screen, the reconstructed Joaquin advocated for “a mix of stronger gun control laws, mental health support and community engagement,” chatted about Remember the Titans and Star Wars, and articulated political positions in a measured, on-message way. His father, Manuel Oliver, explained that bringing “AI Joaquin to life” would “create more impact,” and that the model drew on what his son had written and posted online along with information from the wider internet.
It is impossible to watch this without feeling the moral weight on both sides. These are grieving parents using every tool available to keep their murdered child present and to fight for a cause they believe might have saved him. To call it exploitative would be obscene. And yet the format produced exactly the unease the paper predicts. A teenager who never reached an adult political consciousness was given polished adult opinions, in his own face and voice, for an audience that could not interrogate their provenance. The avatar said reasonable things. That is part of the problem. Because the same machinery, in other hands, could just as easily have made him say the opposite, and the audience would have had no way of knowing which version, if either, reflected the boy who died.
This is the scenario the law is least equipped to handle. The harm, if there is harm, is not financial. It is dignitary and informational, a wrong done to the integrity of a person's identity and to the public's ability to know what a real human being actually thought. Existing frameworks, built around property and market value, have almost no vocabulary for it. The deceased cannot be defamed in most legal systems, because the dead have no reputation to protect in law. The family's distress may not rise to any recognised cause of action. And the company that built the model can point, accurately, to the fact that everyone involved consented, that the parents asked for it, that no statute was broken. Everything was permitted. Nothing was governed.
The Regulators Who Are Not Quite in Charge
If the law of the dead is full of holes, the regulation of AI is full of doors that do not quite open onto this room.
The European Union's AI Act, the most ambitious attempt yet to govern these systems, reaches deadbots only at the margins. Its transparency obligations, which come into force on 2 August 2026, require that people be told when they are interacting with an AI system unless that fact is already obvious, and that synthetic audio, images, video and text be machine-readably marked as artificially generated. That is genuinely useful. It means a well-behaved deadbot should, in Europe, carry a label. But labelling is a thin shield. It tells you that the voice consoling you is a machine. It says nothing about whether the machine should exist, who may build one of whom, what it is allowed to say, or what happens when it causes psychological harm to someone already in the most vulnerable state a human can occupy. The paper makes the sharp observation that formal transparency compliance may even operate as a liability shield, a box ticked that lets relational and psychological harm proceed unimpeded. We told you it was AI. The rest is on you.
The structural problem the authors identify is what they call category indeterminacy. Modern regulation works by sorting things into tiers, high-risk and low-risk, this kind of system and that kind, and a deadbot resists the sorting. Embedded inside a larger platform, it can be treated as user-generated content, which in the UK's Online Safety Act regime, for instance, can leave the underlying model architecture outside the scope of oversight entirely. Considered as a conversational agent, it attracts only light-touch transparency duties. Considered as a processor of personal data, it escapes through Recital 27's exemption for the dead. Each regulatory regime, looking at the deadbot, sees a different object, and concludes that some other regime is responsible. Liability, the paper notes, is rarely obvious, dispersed as it is across platform, developer and user. When everyone is partly responsible, the practical result is that no one is.
Academic and ethical bodies have been clearer than legislators about what good practice might look like. In 2024, the Cambridge researchers Tomasz Hollanek and Katarzyna Nowaczyk-Basińska published a set of design scenarios that have since become reference points. One, called MaNana, imagines a service that builds a grandmother deadbot without the grandmother's consent, comforts the bereaved grandchild for free, then, once the trial expires, begins suggesting takeaway orders in the dead woman's voice. Another, Paren't, imagines a terminally ill mother leaving a deadbot to help her eight-year-old son grieve, raising the question of what it does to a child to be parented by a simulation. The researchers called for safeguards against unwanted digital “hauntings,” for design protocols that prevent deadbots being used for advertising or maintaining a social media presence, and for prompts that force the living to confront the dignity of the dead before resurrecting them, questions as simple as whether they ever discussed with the person how they wished to be remembered. These are sensible proposals. They are also entirely voluntary. Nothing compels a company to adopt them, and the commercial incentive, as the takeaway-advertising scenario suggests, runs the other way.
The Specific Danger of the Present Tense
There is one more thread, and it belongs to the clinicians rather than the lawyers, because it explains why the indeterminacy is not merely an intellectual scandal but a potential source of real harm.
Between roughly seven and ten per cent of bereaved adults develop what is now formally recognised in the DSM-5-TR as prolonged grief disorder, a condition marked by persistent, disabling yearning and an inability to re-engage with ordinary life. For that population, a technology engineered to simulate the continued presence of the dead carries a specific clinical risk, and it is a risk that follows directly from the design. A deadbot, by its nature, operates in the present tense. It does not say your mother loved you. It says, in her voice, I love you, now, today, in response to the message you just sent. It is built to sustain interaction, because sustained interaction is the business model, and it offers the bereaved a relationship that never ends, never grows impatient, never insists on the one thing that mourning requires, which is the acknowledgement that the person is gone.
No US federal law, as of the spring of 2026, sets a psychological safety standard for these products. None of them is subject to the kind of emotional-harm regulation that governs, say, a medical device or a pharmaceutical. A grieving person can buy, with a credit card, a system that may quietly entrench the very condition that makes them most in need of protection, and there is no regulator whose job it is to check. The social indeterminacy the paper describes, the absence of any agreed line between healing and harm, is not a gap that will be filled by better engineering. It is a gap that can only be filled by a decision about responsibility, and so far no institution has volunteered to make it.
Whose Job Is It to Decide
Which returns us to the question underneath all the others. When a commercial product can reconstruct a human being after death, speak in their voice, sustain a relationship with their grieving family, and potentially say things they would have despised, and when there is no clear legal basis for who owns it, who profits, or who answers when it goes wrong, whose responsibility is it to decide what the dead are allowed to become.
The honest answer the research points toward is that no single party can hold it, and the current arrangement, in which the question is answered implicitly by whoever happens to be in the room, is the worst of all options. The companies cannot be the deciders, because their incentive is engagement and their solvency is contingent and their terms of service can be rewritten or auctioned. The families cannot be the sole deciders, because their grief, however legitimate, can author a version of the dead that the dead never agreed to, as the Pelkey and Oliver cases gently demonstrate. The deceased cannot be the deciders, because they are dead, and because the consent they could have given while alive can never have anticipated what a generative model would one day make of them. And the regulators are not yet the deciders, because each of them, peering at the deadbot through the lens of their particular statute, sees a problem that belongs to someone else.
The paper's contribution is to refuse the comforting narrative that this is a temporary lag, a matter of waiting for legislation to mature. Indeterminacy across all five dimensions, it argues, is not a phase. It is the nature of the thing. A perfectly faithful deadbot is technically impossible, because the medium is nondeterministic. A culturally universal standard for healthy grief does not exist, because grief is not universal. The metaphysics of what a simulation of a person even is remains genuinely unresolved. And the law that might govern the dead was built around the living and dissolves at the moment of death. You cannot legislate your way out of a fog by passing a single statute, because the fog is in the categories themselves.
What follows from that is not paralysis but a different kind of seriousness. It means treating the resurrection of the dead as something that requires affirmative justification rather than mere permission, the way we treat other irreversible acts performed on people who cannot speak for themselves. It means building the dignity of the deceased into the design from the first prompt, as the Cambridge researchers urge, rather than bolting on a transparency label at the end. It means data protection regimes that do not simply switch off at the graveside, succession frameworks that treat a digital self as something more than an asset in a bankruptcy, and a settled decision about which regulator owns the harm rather than a polite consensus that it must be somebody. Above all it means accepting that the most consequential choices here, what a dead person may be made to say, to whom, for how long, and for whose benefit, are being made right now, every day, in the absence of anyone with clear authority to make them.
Christopher Pelkey's simulation forgave his killer, and a courtroom found it moving, and perhaps it was. But the man himself was four years dead and could neither grant nor withhold that grace. Joaquin Oliver's avatar argued for gun reform with a fluency the murdered teenager never lived to develop, and his parents found in it a kind of impact, and perhaps they were right. The unsettling truth in both cases is the same. The dead are already being remade, in their own voices, by whoever has the data, the software and the motive, and the question of what they are allowed to become has been answered by default, by everyone and therefore by no one. Deciding it on purpose, before the industry decides it for us, is the unfinished work the law has barely begun.
