There is a right and a wrong way to build an AI of a person. The whole difference is consent.

There is a right and a wrong way to build an AI that sounds like someone you love, and the entire difference is consent, given while the person is alive. An AI they built from their own memories, agreeing to be remembered this way, is a chosen likeness. One assembled from the dead who never agreed is taken from them. Consent is the line, and it can only be drawn from one side of it.

On what would have been Joaquin Oliver's twenty-fifth birthday, a journalist sat down to interview him. Joaquin had been dead for seven years. He was seventeen when he was killed in the 2018 shooting at Marjory Stoneman Douglas High School in Parkland, Florida, one of seventeen people murdered that day. His parents have spent the years since campaigning to end gun violence, and they built an AI version of their son to carry his voice into a fight he can no longer join. On 4 August 2025 the former CNN anchor Jim Acosta published a recorded conversation with that AI, and the reaction was immediate and furious (TheWrap, 5 August 2025).

I want to be careful about why it was furious, because the easy reasons are the wrong ones, and the wrong reasons will lead us to ban the wrong things. It was not furious because the technology was crude, or because grief is private, or because the parents were anything other than heartbroken people doing what heartbroken people have always done, which is refuse to let a child go quietly. It was furious because a seventeen-year-old cannot agree to become a chatbot, and he was made into one anyway, by people who loved him, for a cause he might well have shared. Love was present. Grief was present. The one thing that was structurally, permanently absent was Joaquin. This essay is about that absence, because it is the whole argument.

What is the actual difference between a right and a wrong way to build an AI of a person?

Start with two AIs that behave identically. Both speak in the person's cadence. Both know the family stories. Both can be asked a question and answer the way the person would have answered. From the outside, at the interface, they are indistinguishable. And yet one of them is an act of authorship and the other is an act of appropriation, and no amount of studying the output will ever tell you which is which. The difference is not in the model. It is in the history of how the model came to exist.

The cyberpsychologist Elaine Kasket, a visiting professor at the University of Bath's Centre for Death and Society, puts the whole field's central question in nine words: "Did the deceased agree, explicitly or implicitly, to become a chatbot?" (Elaine Kasket, "What Are Grief Bots?"). She warns that without clear policies, "we risk creating a digital afterlife in which the dead are remixed, monetised, or reinvented without their agency or the informed consent of their loved ones." Read that phrase slowly. Remixed. Monetised. Reinvented. Those are things done to an object. They are not things that happen to a person who is still in the room.

So here is the distinction stated as plainly as I can state it. A consent-first AI of a person is built forward, by the living, as a decision. The person sits down, chooses what to record, chooses what to leave out, and agrees, in advance and in writing, to be remembered this way after they die. A scraped AI is built backward, from the outside, after death, from text messages and voicemails and social feeds the person left behind for other purposes, by someone deciding on their behalf that this is what they would have wanted. One is a will. The other is a guess dressed as a tribute.

The output can be identical. The ethics never are. One was authored by the person. The other was authored about them.

The reason this matters more every year is that the guessing is getting easier and the guessing is getting cheaper. The tools that used to require a research lab now require an afternoon, and the number of people offering to do the guessing for you is climbing fast.

The ethics of AI personas of the dead, by the numbers (as of July 2026)

The debate is not hypothetical, and it is not fringe. It is a live industry with a growing academic literature holding it to account. Here is the honest state of it, sourced.

• More than half a dozen platforms now offer to recreate the dead as interactive AI, according to a news feature in one of the world's leading scientific journals (Nature, September 2025). Proponents say it comforts mourners; skeptics say it complicates grieving. Both can be true.

• A single peer-reviewed study analyzed six of these services by name: Seance AI, Eternos, You Only Virtual, HereAfter AI, Project December, and re;memory, examining the assumptions they make about what "normal" grief is allowed to look like (ScienceDirect, January 2026).

• The most-cited ethics paper in the field, from Cambridge researchers, lays out three distinct scenarios for how these systems get built, from a person donating their own data to a company reconstructing someone from what they left behind, and recommends mutual consent, dignity protections, and dignified ways to retire a bot (Hollanek and Nowaczyk-Basińska, Philosophy & Technology, 9 May 2024).

• The practice is older than the current AI wave. Project December, one of the first services to let people build chatbots of the dead, launched in September 2020 on an early large language model, went viral after a man used it to revive a simulation of his late fiancee, and had its underlying model access restricted by its provider soon after (The Register, 8 September 2021).

• Mainstream science journalism now treats the question as a standing beat rather than a curiosity: a November 2025 feature weighed whether these systems help mourners heal, canvassing therapists and philosophers on both sides (Scientific American, 18 November 2025).

Notice what that literature is converging on. Not "ban it." Not "build it freely." The recurring word, in paper after paper, is consent. The field has already found the load-bearing wall. The disagreement is only about whether anyone will build on it.

Why does consent from the living person change everything?

Because the dead cannot object, and a right that cannot be exercised is not a right. This is the asymmetry that sits under every hard case. When you build an AI of a living person without asking, they can find out, be appalled, and make you stop. When you build one of a dead person without having asked, there is no one left who can say no. Their silence is not agreement. Their silence is the whole problem.

A dedicated ethics paper on grief bots argues that consent has to cover three separate acts, not one: consent to the creation of the persona, consent to its ongoing use, and consent to its deletion, plus a specific consent to use the person's likeness and voice (ScienceDirect, 2026). Think about how many of those the dead cannot give. A person recreated after death consented to none of it. Not the making, not the running, not the ending, not the face, not the voice. Every one of those decisions was made for them, by someone standing outside their life, reading it like a book they never agreed to have published.

The Cambridge researchers frame the harm in terms of dignity. "It's important to prioritize the dignity of the deceased," says Katarzyna Nowaczyk-Basinska, "and ensure that this isn't encroached on by financial motives of digital afterlife services" (University of Cambridge, 9 May 2024, reported via SciTechDaily). Dignity is exactly the right word, because dignity is the thing consent protects. A person who chose to be remembered a certain way keeps their dignity even in a machine, because the machine is carrying out their wishes. A person reconstructed against, or simply without, their will has had their dignity decided by a committee they were not invited to.

The dead cannot consent. That is not a technicality to route around. It is the reason the whole question exists.

This is why the Parkland case is the hard case and not the clean one. Joaquin's parents are not villains, and I will not pretend otherwise to make a tidier argument. They acted from grief and from conviction. But a persona of a minor who never agreed, built and voiced after his death to advocate a position, cannot draw its authority from Joaquin, because Joaquin never granted any. It can only draw its authority from his parents, and the moment we admit that, the "interview" becomes what it actually is: his parents' message, in his stolen voice. That may be a message worth hearing. It is still not his.

But isn't a griefbot that genuinely comforts a mourner a good thing?

This is the strongest objection to everything I have written, and I want to argue it at full strength rather than knock down a weak version. Grief is not a problem to be solved, but it is a wound, and comfort is not nothing. Therapists interviewed across the recent coverage describe patients who found real solace in one more conversation, one more chance to say the thing that went unsaid (Scientific American, 18 November 2025). Some researchers studying older and isolated mourners have found the interactions eased loneliness. If a widow spends twenty minutes hearing a familiar way of speaking and stands up lighter, who exactly am I, or any ethicist, to tell her she did something wrong. That is a serious question and it deserves a serious answer, not a slogan.

Here is the honest concession. The comfort is real, and a blanket condemnation of every griefbot is a comfortable position held by people who are not the ones grieving. Refusing that comfort on principle can be its own kind of cruelty.

And yet the comfort of the living does not, and cannot, answer the consent of the person being recreated, because they are two different people with two different claims. The mourner's need is real. The dead person's absence of agreement is also real. When those collide, "but it helps me" is an argument about the mourner's welfare, and the ethical objection was never about the mourner's welfare. It was about whether we get to conscript someone into a performance they never auditioned for, however much their performance would soothe us. A too-perfect replica, built to meet our need, has quietly replaced the person with the need itself. The comfort is genuine. It is also, when there was no consent, comfort purchased with someone else's dignity, and the person paying does not get a vote because they are dead.

There is a further cost, one the mourners themselves eventually pay. Some clinicians and writers now worry that an always-available bot lets a grieving person keep the relationship external and running instead of doing the harder internal work of letting the dead become memory (Nature, September 2025). I hold that worry loosely, because the research is young and grief is not one thing. But it points at the same fault line from the other direction. A persona the person built for you, and finished, and sealed, is a completed gift you can return to. A persona you keep generating, forever, on demand, from a model that will answer anything you ask, is a door you have wedged open onto a room the person already left.

What about the people who have to receive the message?

Consent, it turns out, has two sides, and the field's sharpest insight is that we keep forgetting the second one. It is not only the recreated person who must agree. It is also the person on the other end of the conversation, the one the persona will speak to.

Tomasz Hollanek, one of the Cambridge authors, puts it directly: "It is vital that digital afterlife services consider the rights and consent not just of those they recreate, but those who will have to interact with the simulations" (University of Cambridge, 9 May 2024). He warns of what he calls "unwanted digital hauntings," a phrase worth keeping. Picture the persona a grieving parent builds of a child, then wills to a sibling who did not want it, who now receives messages from a dead brother they were still learning to grieve. Picture a service that keeps a persona active, and keeps its subscription live, by sending occasional unprompted messages, because engagement is the business model and a silent bot does not renew.

This is where good intentions and bad incentives come apart. A persona built with the recipient in mind, handed over gently, at a moment of the living recipient's choosing, honors both consents. A persona pushed at people, autoplaying in a moment of raw grief, extracting attention because a spreadsheet needs it to, honors neither. Nothing about a legacy should ever arrive as an ambush. The rule I hold, and I will say plainly that no academic paper states it in these words so I offer it as our design principle rather than a cited finding, is this: nothing in a grief context should ever autoplay, and no message from the dead should ever arrive uninvited. Every encounter should be a chosen tap. A gift is offered. Only a haunting is imposed.

Consent has two sides. The person recreated has to agree to speak. The person listening has to agree to hear.

Put the two consents together and you have the whole ethical test, and it is a test you can actually apply. Did the person agree, while alive, to be remembered this way? And is the person receiving the persona free to accept it, on their own terms, in their own time, or walk away? If both answers are yes, you are looking at a gift. If either is no, you are looking at something taken from one person and pressed on another, no matter how much love was in the room when it was made.

The standard a consent-first company owes you

I build in this industry, so I am going to say what I think the standard is, knowing that I am handing you the exact yardstick to measure my own company against. That is the point. A standard you cannot use to hold the person stating it accountable is not a standard, it is marketing.

Here is what I believe any serious, consent-first company owes the person it recreates. The persona must be built by the living person, on purpose, from what they chose to include, never scraped from the dead and never assembled from the outside. Consent must be explicit, documented, and specific: to the creation, to the use, to the voice, and to what happens after death. The person's wishes, once set, must be honored and not quietly rewritten later by a family member, an heir, or the company itself. And the whole thing must be built to be received as a gift and not imposed as a haunting: chosen, never autoplayed, never used to farm a grieving person's attention.

That standard is why Afterlife.ai® is built the way it is, and why we say no to things that would be easier to say yes to. Your Persona is created only by you, while you are alive, from memories you choose to add, one decision at a time. We do not build a Persona of anyone from the outside after they have died, because there is no one left to consent, and a Persona without consent is exactly the thing this essay argues against. Voice is preserved with professional voice technology and only ever from your own voice, with your explicit agreement that it may speak after you are gone, an agreement made while you can still give or refuse it. And Executor Lock™ exists to answer the question the scraped services cannot: at a moment you choose, the record you built is sealed as a perfect snapshot of who you were. Every conversation after that draws only on that sealed record, and your Executor governs access but cannot rewrite who you were. Your consent, frozen. Your dignity, held.

I will state the limits with the same plainness, because a company that only lists what it can do is selling something. We cannot recreate a person who did not build themselves. We cannot recover a voice that was never recorded with consent. We cannot promise that any AI, ours included, is a substitute for a person, and we will not pretend that it is; a Persona is a living likeness the person authored, not the person returned. If you want the fuller version of why building your own, while you are alive, is the only version of this that is yours to give, we made the case for building an AI of yourself separately. And if you want to see how the consent-first approach differs, point by point, from the griefbots this essay has been circling, that comparison is its own page.

What should you do about this while you are alive?

The uncomfortable answer is that the only person who can build the right kind of AI of you is you, and the only time you can do it is now, while you can still choose what goes in and agree to how it is used. Every hard case in this essay exists because someone waited until the person could no longer consent, and then loved them so much they went ahead anyway.

So the practical move is not to fear the technology or to swear it off. It is to take the decision out of your family's hands and put it in your own, where consent lives. Decide, while you are healthy, whether you want to be remembered this way at all, because the right to say no is as much a part of consent as the right to say yes. If you do, build it yourself, deliberately, from what you actually want carried forward, and set the terms for after you are gone rather than leaving them to be guessed. If you are weighing whether the voice side of this is safe to do at all, we wrote plainly about whether voice cloning is safe and where the real risks sit. And because these decisions do not stop at persona-building, it is worth understanding who actually owns your digital afterlife once you are no longer here to manage it.

The Parkland case will keep being cited, and it should be, because it is the honest edge of this whole question: the place where love and consent came apart, and love went ahead alone. The lesson is not that his parents did something monstrous. The lesson is that the only person who could have made that AI rightful was Joaquin, and the only time he could have done it was while he was alive. That window does not reopen. It is open for you right now, and only right now, which is the one thing every ethicist, every court, and every honest builder in this field agrees on: the decision belongs to the person, and the person has to be alive to make it.

Build the version of you that you consent to. Then you never leave the question to the people who loved you too much to ask.

Frequently asked questions

What makes an AI persona of a dead person ethical or unethical?

The dividing line is consent from the person being recreated, given while they were alive. An AI a living person built themselves, choosing what to include and agreeing in advance to be remembered this way, is ethical because its authority comes from them. An AI scraped from a dead person's messages and voice after their death, however loving the intent, cannot draw authority from someone who never agreed. The behavior of the two can look identical; the ethics never are. Consent, as the Cambridge researchers and Elaine Kasket both argue, is the field's load-bearing principle.

Is it wrong for grieving families to use griefbots for comfort?

Not simply, no. The comfort is real, and blanket condemnation is easy for people who are not the ones grieving. The ethical objection is narrower and sharper: comfort to the living does not answer consent from the person being recreated, because those are two different people with two different claims. A persona the person built and sealed themselves offers comfort without taking anyone's dignity. One reconstructed from a dead person who never agreed offers the same comfort at someone else's expense, and that person cannot object because they are gone.

Can you build an AI of someone who has already died?

Technically yes, and more than half a dozen services offer to (Nature, September 2025). Ethically, this is the practice this essay argues against, because a person who has died can consent to none of it: not the creation, not the voice, not the ongoing use, not the ending. That is why Afterlife AI™ only lets you build a Persona of yourself, while you are alive. The right kind of AI of you can only be authored by you, which is also the reason to start now rather than leave it to the people who will one day wish they could ask.

Who has to consent, just the person being recreated?

Both sides. The person recreated must agree, while alive, to be remembered this way. But the Cambridge researchers stress a second consent that is easy to forget: the person who will interact with the persona has to be free to accept it on their own terms, or decline. Tomasz Hollanek warns of "unwanted digital hauntings," personas pushed onto people who never asked for them. A legacy handed over gently, at a moment the recipient chooses, honors both consents. One that autoplays or arrives uninvited honors neither.

How is Afterlife AI™ different from a griefbot?

A griefbot is typically built after death, from the outside, from data the person left behind for other purposes. A Persona on Afterlife AI™ is built by you, while you are alive, from memories you deliberately choose, with your explicit consent to how your voice and likeness may be used after you are gone. Executor Lock™ then seals who you are at a moment you choose, so no one can rewrite you later. We do not build Personas of the dead from the outside. The full comparison lives on our page on the ethical alternative to griefbots, and the plans are on pricing.