The Lines We Will Not Cross, Beginning With an AI of a Person Who Has Died

Technically, yes, and the job gets easier every quarter. Three years of messages, a few hours of voice and one wedding video are more than a system needs to produce something that answers in a dead man's cadence. The honest answer is the other half: a system can produce the likeness, and nobody alive can produce the permission.

Here is what we write back. We build for the living. The person being preserved decides what is recorded and who may reach the result, while they can still change their mind. Your father did not get that chance, so we cannot give you what you are asking for.

Then the two things that matter more than our product. Keep the files, in two places, because the recordings you already hold are the legacy. And if the folder weighs more than a folder should, grief support that involves a human being exists, and asking for it is not a failure of love.

I said the same to Ricardo F. Colmenero of EL MUNDO on 30 August 2026, when he asked whether any line existed that I would not cross for money: "Como empresa, eso es algo que no hacemos. Es una línea ética muy importante." As a company, that is something we do not do, and the line matters.

We can build the likeness. We cannot build the permission.

No, unless the person said yes while they could still say no. That is the entire test, and everything else in this industry is detail.

The word people search for is "deadbot". A deadbot is a conversational system assembled from a dead person's data by somebody else, after that person lost the ability to refuse. Tomasz Hollanek and Katarzyna Nowaczyk-Basińska of the University of Cambridge's Leverhulme Centre for the Future of Intelligence published the most careful ethical treatment of these systems in Philosophy & Technology in 2024, and their recommendations to providers included "restricting access to such services to adult users only" and "adhering to the principle of mutual consent of both data donors and service interactants".

Mutual is the word carrying the weight. Consent, in this industry, is a living person's recorded decision about what may be made from their words and who may reach the result after they die. A bereaved family can supply the data and the longing. They cannot supply the consent, because the only person who could has died. Nowaczyk-Basińska put the commercial risk plainly: "It's important to prioritise the dignity of the deceased, and ensure that this isn't encroached on by financial motives of digital afterlife services." I read that sentence as a description of my job.

There is one exception. If a person wrote down, while alive, that their recordings may be used to build a Persona of them after death, in a will or a signed letter, the consent exists and the door is open. What a clause like that needs to say is a page of its own. In practice the clause almost never exists, so the answer to a grieving reader is still, almost always, no.

The dead cannot consent, and the grieving cannot consent for them.

Six things, written down before the first customer asked, because the first customer was going to be crying, and a rule invented in front of a crying person bends.

Video. Where video exists at Afterlife AI, the face is made only from the person's own recordings, with their consent, behind quality gates we set deliberately high. A wrong pause in a voice sounds like a bad phone line. A mouth that lands a frame late on a mother's name is a wound.

Age. Nobody under 18, or under the age of majority where they live, uses the product. A child grieving a parent is the most sympathetic case in this field and precisely the one that must wait, because consent given by a minor to a machine is not consent.

The lock. Executor Lock is the process that freezes a Persona as a perfect snapshot at verified death, so that nothing is added to the personality, no retraining occurs and no drift creeps in. Drift, in an AI system, is the slow movement of a model away from the person the data described and toward a character the model finds easier to play. Executor Lock exists so the inferences have a ceiling and the facts have a floor.

Grief. No streaks, no autoplay in grief contexts, no notification engineered to bring a widow back at eleven at night. When the product detects a person in distress, the Persona stops, the person is referred to a local crisis line with the right number for their country, and the conversation ends.

Transparency. A Persona is a rigorous replica of a living person, built from their own words and voice with their consent, and everyone who speaks with one is told exactly that. A product that tells you only the first half is selling a séance.

One ugly thing, stated once. During development someone uploaded a photograph of a person they loved and asked the system to remove her clothes. The request was refused, and the guardrail that catches it exists because of that afternoon.

A rule that is not written down is a mood.

Every line in that table turns away a paying customer, and the sum of them is the only reason anyone should trust us with a voice.

The arithmetic is not subtle. The people who most want to buy in this category are bereaved, and the request we refuse most often is the one they would pay the most for. A company that builds the dead has a bigger market than a company that builds the living; the second company simply has customers who can still say no. The other lines cost in the same currency, the lock most of all, because "your father keeps learning" is the most lucrative sentence in this field and we will never say it.

Then there is the cost that has nothing to do with rules. HereAfter AI, the interview app co-founded by the journalist James Vlahos, replaced its product pages in July 2026 with a notice that began "unfortunately HereAfter is shutting down" and asked families to email support to retrieve their recordings; on 5 September 2026 the homepage returned a not-found page. What the HereAfter AI shutdown means for the families who recorded there sets out how to get the files back. HereAfter did nothing wrong. In this category a company going quiet is a second death, and Colmenero asked me exactly that. The risk exists. Your memories and conversations are exportable today, the lock guarantees the Persona cannot be altered by anybody, including us, and one day I expect something like a piece of hardware that holds a personality outside any company. That last part is not built.

So we refuse the bereaved and hold back the features that demo best, and in exchange a living person can hand us the most intimate recording they will ever make and know precisely what will and will not be done with it. That may sound bad for short-term revenue. Long term, it is the only business in this category that survives contact with its own customers.

Trust is the product. The rest is software.

Start building your legacy Free build, 50 memories, no card required.

They each patented the thing we refuse, and each says there are no plans to build it.

On 30 December 2025 the United States Patent and Trademark Office granted Meta Platforms Technologies patent US12513102B2, "Simulation of a user of a social networking system using a language model", filed on 29 November 2023 and listing Andrew Garrod Bosworth, Meta's chief technology officer, as inventor. The abstract says the model "may be used for simulating a user that is deceased", so that "other users can continue to experience the presence of the target user in spite of the fact that that the target user is deceased". A Meta spokesperson told Fortune in March 2026: "We have no plans to move forward with this example." In the same piece the cyberpsychologist Elaine Kasket observed that the patent "suggests that user death is like an engagement problem".

Microsoft got there first. In December 2020 the company was granted a patent, applied for in April 2017, for a conversational chatbot modelled on a "past or present entity … such as a friend, a relative, an acquaintance, a celebrity, a fictional character, a historical figure". Tim O'Brien, then Microsoft's general manager of AI programs, wrote in January 2021 that he had "confirmed that there's no plan for this".

A patent is a legal right to a method, held whether or not the method is ever built. I take both companies at their word, and I notice which way the incentives point: a platform's revenue comes from attention, a dead user stops producing attention, and the Meta patent keeps the attention flowing from data collected for a different purpose while the person was alive. Jason England of Tom's Guide drew the distinction in February 2026: services like ours are "opt-in and focused on legacy", and the patent is neither. When Colmenero asked which large company I would stop from using this technology, I named Meta. The Facebook death patent, read line by line goes through the document.

A patent is a drawer, and drawers open.

Yes. Every company in this field publishes principles, and a principle is worth exactly what the mechanism behind it can bear. Test mine.

Start with the request: write to us and ask us to build a person who has died, and read the reply, which will not change if you mention a budget. Then test the gate. Ask who can open a Persona after a death, what evidence they must file, how many days that takes, whether the builder can stop the release, and whether every other named person is told. Executor Lock's seven-day, evidence-verified process publishes all five answers, and a company that cannot answer those five in a sentence each has decorative consent. Then read how we govern identity, from consent capture to release and why trust is our foundation, and if you find a gap between either page and the table in this essay, write to us: we will fix the page or say why we have not.

A company can publish its mechanism and still change its mind after the funding round or the acquisition. True, and the reply is structural: a locked Persona cannot be altered by the company holding the record, the export exists so a family is never hostage to our survival, and the refusals above are written into the product rather than into a page somebody can edit. A rule enforced by code outlives the person who wrote it.

Test the mechanism, not the manifesto.

The fact that closes the door on rebuilding the dead opens one for the living. Your data is already held by companies whose patents describe simulating you after you die, and your voice is in voicemails and old videos on other people's phones. The question is no longer whether a likeness of you can be made without your permission. It can. The question is whether the authoritative version of you will be the one you made on purpose, with your consent recorded and your rules attached, or the one somebody else assembles later.

I told EL MUNDO that a personality you create and legally validate as yours, with your consent, is what prevents someone else from creating an unauthorised one: almost a defence mechanism for controlling your own image and identity. A Persona built by you is a dated, consented, locked record of who you were and of what you wanted done with your words, and a will only has to point at that record.

That is the upgraded thesis. A company in this industry is defined by what it refuses. A person in this decade may come to be defined by what they consented to before anybody else decided for them.

To the reader who wrote to us about a father: keep the files, in two places. They are already a legacy. To the reader whose father is alive and in the next room: the free build is 50 memories, no card, and never expires, and the most consent-first act in this industry is handing the microphone to the person whose life it is, while they can still decide what to say.

We refuse to build the dead so that we can be trusted with the living.

Is it ethical to make an AI of a dead person?

Not unless that person consented while alive, clearly and in writing, to being recreated after death. The dead cannot refuse, and a grieving family cannot refuse on their behalf. University of Cambridge researchers who studied these systems in 2024 recommended mutual consent from the person whose data is used and the person who talks to the result, and the first of those can only be given while alive. Afterlife AI builds Personas only for living people and declines every request to recreate someone who has died.

Can I upload recordings of my dead father to Afterlife AI?

No, and we say so to every family who asks, which is the request we receive most often. Afterlife AI builds a Persona only for a living person, who decides while alive what is recorded and who may reach the result after death. Your father did not have that chance, so we cannot make a Persona of him. The one exception is explicit written consent he left himself, for example a clause in his will. If that exists, write to us. Either way, keep the recordings safe.

What if the person said in their will that an AI of them could be made?

Then consent exists, and that is the single exception Afterlife AI recognises to its rule against recreating people who have died. The clause needs to be the person's own, written while alive, and clear about what data may be used and for what. A vague line about "my digital assets" is not enough. Such clauses are still rare, which is why the answer to most bereaved families is still no. If you are alive and want this for yourself, build and lock your own Persona now, so the will only has to point at something real.

Does Afterlife AI change a Persona after someone dies?

No. At verified death, Executor Lock freezes the Persona as a perfect snapshot of everything the person recorded: nothing is added to the personality, no retraining happens and no drift is allowed. Family may still hear the Persona and ask questions, but the answers come only from what the person actually said while alive. This is the opposite of a system that "keeps learning" after death, and we refuse that design on purpose, because a personality that keeps moving stops being the person and becomes a character.

What does Afterlife AI do if someone is in crisis while talking to a Persona?

The conversation stops. When the product detects a person in distress, the Persona refers the person to a local health service or crisis line with the correct number for their country, and ends the conversation. We do not treat grief as an engagement metric, so there are no streaks, no autoplay in grief contexts and no nudges designed to pull a grieving person back at night. If you are struggling now, please contact a crisis line where you live, or a person who can sit with you.

Why does Afterlife AI not let people under 18 use it?

Because grief and consent both need an adult. A child grieving a parent is the most sympathetic case in this field and exactly the one that must wait, because consent given by a minor to a machine is not consent. Afterlife AI does not interact with anyone under 18 or under the age of majority where they live, and the Cambridge researchers made the same recommendation in 2024. A parent who wants a child to hear them one day builds the Persona now and names who may hear them, and when.

• Whether it is ethical to build Personas of people who have died: the full argument from the side of the person who cannot consent.

• Can Facebook recreate dead people: what a platform can do with the data it already holds.

• The hologram at the Christmas table: the EL MUNDO headline, and what has to be true first.

• Consent for an AI of me in my will: the only exception to the first line, in detail.

• Talk to an AI version of someone who died: what exists, and what to do with recordings you hold.

• 2wai vs Afterlife AI: the company that crossed the first line, and what the backlash taught.

• Uare.ai vs Afterlife AI: a rival that shares the fear and answers it the other way.

• Afterlife AI vs ChatGPT: the assistant serves whoever is typing; the fourth line says a legacy must not.

Corrections: tell us if anything here is wrong or out of date.

• EL MUNDO, Ricardo F. Colmenero, Chris Williams, CEO de Afterlife: "En cinco u ocho años podrás hablar con el holograma de un familiar fallecido en la mesa de Navidad", PAPEL, 30 August 2026.

• Google Patents, US12513102B2, Simulation of a user of a social networking system using a language model, Meta Platforms Technologies LLC, filed 29 November 2023, granted 30 December 2025.

• Fortune, Jacqueline Munis, on Meta's patent for an AI model that simulates deceased users, 3 March 2026.

• Tom's Guide, Jason England, My ghost is not for sale: The case against Meta's AI Afterlife patent, 19 February 2026.

• CNN Business, Clare Duffy, Microsoft patented a chatbot that would let you talk to dead people, 27 January 2021.

• University of Cambridge Apollo repository, Hollanek and Nowaczyk-Basińska, Griefbots, Deadbots, Postmortem Avatars, Philosophy & Technology, 2024.

• University of Cambridge, Call for safeguards to prevent unwanted 'hauntings' by AI chatbots of dead loved ones, 9 May 2024.

• HereAfter AI homepage, a not-found page when checked on 5 September 2026.

Chris Williams is the founder and CEO of Afterlife AI, a consumer brand of Idy Pty Ltd in Sydney, and the architect of Executor Lock, the process that locks a Persona as a perfect snapshot at verified death. He is a father of four. He has been interviewed on consent-first digital legacy by EL MUNDO, Channel 10 News, The Daily Telegraph and ABC radio. Before Afterlife AI he founded Natural Solar, an Australian solar energy company.