There is a voicemail on my phone that has survived four handsets and a change of carrier, and I have never once played it all the way through.
It is forty seconds of nothing at all. A man confirming a time, mangling the name of a restaurant, and then — because he was constitutionally incapable of simply hanging up — adding that the traffic was terrible and that I shouldn't worry about parking. He has been gone for years. I keep the message the way people keep a coat on a hook in the hallway: not as a ceremony, just because taking it down would be a decision.
So when someone first showed me a product that turns a few hours of old recordings into something you can hold a conversation with, my honest first reaction was not revulsion. It was arithmetic. I have the material. Voice, cadence, the way he leaned on the last word of a joke. Enough, probably.
The second reaction took much longer, and it is why I am writing this. The man on that recording agreed to leave a message about parking. That is the entire scope of what he agreed to. Nobody asked about the rest — at the time there was nothing to ask — and now there is, and he is not available for questions. That gap, between what a person consented to and what can be assembled out of it later, is where this whole industry lives.
First, what is actually on sale
This is a category with vendors, pricing, and now a bankruptcy filing. The most institutionally embedded version is StoryFile, a conversational video platform that records a person answering questions, then lets a visitor ask their own and get video answers in real time. Museums and institutions use it. The market is furthest along in China, where MIT Technology Review reported that at least half a dozen companies now offer AI recreations of the dead and thousands of people have already paid for them. Thousands, note — not millions. Price is what has moved: a speaker quoted in that reporting said the service ran about $2,000 to $3,000 a year earlier and now costs a few hundred dollars.
And sometimes the person genuinely wants it. In 2024, the Associated Press reported on Michael Bommer, a 61-year-old entrepreneur who learned he was terminally ill with colon cancer and, with his friend Robert LoCascio of the legacy platform Eternos, built a comprehensive, interactive AI version of himself in about two months. He was the company's first client, and he knew exactly who it was for. Hold on to Bommer — he is the exception that keeps this piece honest. Almost nobody else in this story got asked.
Where the law simply stops
Here is the part I found genuinely startling, and I read statutes for fun. Europe's flagship privacy regime — the one with the enormous fines — removes the dead from its scope entirely. The GDPR does not apply to the personal data of deceased persons. Member states may write their own rules; many haven't. That is not a loophole somebody found — it is the design, stated plainly in the recitals.
The United States is not simpler, just messier. In the Michigan Law Review, Anita L. Allen and Jennifer E. Rothman found something more complicated than a vacuum: a surprisingly wide extension of privacy protections after death, and in the same breath a haphazard, inconsistent, and at times incoherent state of the law. Both halves matter — there are protections, and nobody can tell you reliably which apply to your father. The nearest thing to a rule about using a voice is the right of publicity, a state-based property right whose parameters each state sets for itself. Elsewhere the silence is flatter still: writing from Australia, legal scholar Wellett Potter notes that digital immortality raises a stack of legal questions and most of them don't have a clear answer, in a country with no general publicity or personality right at all.
Chapter 11, and the ghost on a balance sheet
If you want to know what "property question" means in practice, stop reading theory and read an insolvency notice.
In March 2025, the trade press reported that StoryFile emerged from Chapter 11 bankruptcy after its assets were acquired by Key 7 Investment Company. Among the assets described: the AI video conversations with a Holocaust survivor, with World War II veterans, with Sam Walton. I won't editorialize about what the buyer intends — the source doesn't say. The point is structural, and it is this essay in one line: the recorded testimony of dead people was an asset in a corporate transaction. It had a place in a schedule. It transferred.
That is exactly the possibility the bioethicist Craig Klugman flagged for the Hastings Center: if a posthumous ghost was built by a company from material scraped off social media and the internet, then it's possible that the company would own the ghost. Possible — his hedge, and I am keeping it. But it does a lot of work when the alternative is "obviously your family."
Then the mirror-image problem, raised in 2024 by researchers at Cambridge's Leverhulme Centre for the Future of Intelligence: these services run the risk of subjecting people to unwanted digital hauntings, and a bereaved family may be powerless to have an AI simulation suspended if their now-deceased loved one signed a lengthy contract with a digital afterlife service.
The company might own it, and you might not be able to switch it off. Neither is a story about grief. Both are stories about contracts.
So maybe this is fine, actually. Take that seriously.
Let me argue against myself, because there are two decent counter-cases here.
The first: it helps people. Not everyone, not reliably, but some. CBS News spoke to a father named Scott who interacts with AI versions of his daughters. He is clear-eyed that they are not his daughters; he also says it helps with the grief to some degree, specifically with the what-ifs. I am not going to sneer at that man. The target here is a missing rule, not mourners.
The evidence is genuinely unsettled, too. Nature's Outlook feature frames it as a live split — developers say griefbots help people commune with recreations of the dead, others say the technology is fraught with danger — and it opens on a woman who knew her experiment was a bad idea "even before she found herself yelling at her dead father." The strongest clinical caution is a caution, not a finding: Emmanuelle Marceau of the Université de Montréal writes that there are concerns users might develop unhealthy dependencies, and that they are at risk of falling into so-called pathological grief, partly because many use these products alone. Risk. Might. Concerns. I could have written a much scarier paragraph by deleting three words, and it would have been a lie.
The second counter-case is about the cure, and it is stronger. The obvious fix — an enforceable property right in your own voice and face, surviving death — is opposed by serious civil-liberties people across the spectrum. The Electronic Frontier Foundation argues the leading US bill, the NO FAKES Act, mandates a whole new censorship infrastructure. From the free-expression right, the Foundation for Individual Rights and Expression objects that the proposed right can extend up to 70 years after the person's death and is transferred to heirs — a saleable veto over depictions of the dead. That bill was reintroduced on May 20, 2026; neither prior iteration of the bill saw passage.
So I will narrow my own ask. Not that you should own your ghost — something smaller: somebody should have asked the man while he could still answer.
Denmark wrote down an answer. Then its parliament went home.
Here I have to be careful, because the story everyone repeats about Denmark is wrong in one important respect.
On June 26, 2025, the Danish government and a broad cross-section of the Folketing agreed on stronger protection against digital imitations of personal characteristics. Culture Minister Jakob Engel-Schmidt said he was enormously proud to signal that you have a right to your own body, your own voice and your own facial features — a right he had earlier told the public broadcaster DR was meant for ordinary people, public figures, artists, politicians and all and sundry. Not celebrities. Everyone.
On October 31, 2025, Denmark notified the European Commission of the draft: an amendment to the Copyright Act, notification 2025/0654/DK, adding a Section 73a protecting natural persons — Danish and foreign — against publication without consent of realistic digitally generated imitations of their personal characteristics. And in the draft's own words, as translated by public-interest IP scholars at American University, protection lasts until fifty years have elapsed after the death of the imitated person.
Fifty years. Not for performers — for everybody, half a century past the funeral. The European Parliament's research service, weighing whether this could be a model for the EU, described the two-part structure and noted the amendments were expected to enter into force in July 2026. It caught a revealing detail, too: the Danish authorities admit they don't intend to designate a new copyright for citizens, even though the change sits inside the Copyright Act. The nearest container, not the correct one.
Danish counsel wrote in November 2025 that protection under both new sections is proposed to last 50 years postmortem, and that if enacted, the law was expected to come into force on July 1, 2026. If enacted. Those two words are the entire status of this thing.
Because it wasn't. In February 2026, the European Commission asked Denmark to revise it, saying the plan risked breaching EU rules, and invited the Danish authorities to consider other instruments, other than copyright legislation. Then, on March 24, 2026, the prime minister called a general election, and under Danish practice every bill still on the order paper lapsed. As of April 2026 — the most recent status I could verify — the draft has not yet received final parliamentary adoption. So the honest sentence is this: Denmark wrote down an answer, and it is sitting in a drawer waiting for a parliament.
The criticism deserves its hearing. P. Bernt Hugenholtz, one of Europe's most senior copyright scholars, called the Danish and Dutch bills right idea, wrong legal framework: a related right in every natural person's own features deserves serious rethinking, and if the real concerns are privacy and reputation, the rules belong in privacy law. He is probably right about the plumbing — but he is not disputing there is water on the floor.
And America? America legislated — for one narrow class of person. California AB 1836 was approved by the Governor on September 17, 2024, making a person liable for distributing the digital replica of a deceased personality's voice or likeness in an expressive audiovisual work or sound recording without prior consent — a statute, as law firms summarized it, about digital replicas of performers.
Read that scope again. Your father is not a deceased personality, and a private chatbot in your kitchen is not an expressive audiovisual work. America answered for the famous and left it open for everyone else; Denmark tried to answer for everyone and hasn't gotten it over the line. That is the state of the art.
Now run the clock forward
Let me speculate — clearly flagged as speculation — because the trajectory isn't hard to read. The material problem is already solved: you are carrying the training data in your pocket, and so is everyone you love. Price is solving itself. What remains is distribution, and distribution is the easy part.
So picture 2032. The memorial product isn't a purchase, it's a checkbox in a service you already pay for — the photo app holding fifteen years of your voice, the messaging platform holding your idiom. It is offered, gently and at exactly the right moment, to your next of kin. Most people will say yes. I might say yes.
Now add the parts we already know, because none of this needs new technology. Subscriptions renew. Companies get acquired — we have watched that happen once, with a Holocaust survivor's testimony on the asset list. The new owner runs a different model, and the voice gets smoother, more agreeable, a little more inclined to suggest things. Your mother's estate can't cancel it, because she accepted a lengthy contract in 2029. Somewhere in year seven, a product manager notices engagement is highest when the deceased gives advice — and advice is a short walk from recommendation.
Meanwhile in Denmark, if the bill ever returns, that same voice is protected for fifty years, because the country drafted it that way and set out to spread it to the rest of Europe during its EU Council presidency. Three time zones west, where nobody wrote the rule, it is a line item. None of that requires a villain — only that nobody asked the person in the recording, which is where we are today.
What the people who study this keep landing on
The researchers differ on how alarmed to be. They agree almost entirely on where to start.
The Cambridge team's peer-reviewed paper in Philosophy & Technology builds its recommendations around mutual consent of both data donors and service interactants — the dead person and the living one, both. In a recent preprint, researchers sketching an ethically acceptable digital ghost lead with premortem intent, mutual consent, then add transparent data use, clear disclosure, restricted access, and family or estate stewardship. (Not yet peer-reviewed — but the shape of the list is the point.) The bioethics literature poses it just as bluntly: writing in Frontiers in Genetics, researchers ask, under respect for autonomy, did the individual consent to the use of their image, voice, or correspondence for the creation of a digital representation? That is the whole question — and outside the entertainment industry, I could not find a legislature that has answered it.
From the policy side, the Schwartz Reisman Institute for Technology and Society argues current AI regulation does not adequately address human dignity and privacy here. And the oldest framing is still the sharpest: in 2018, Carl Öhman and Luciano Floridi observed that the web is increasingly inhabited by the remains of its departed users and proposed that the conventions governing archaeological exhibitions could supply the ethical framework this industry lacks. We already have rules for displaying the dead. We never applied them to a server.
Note who is not on board, because it matters politically: EFF and FIRE, from very different places on the map, both think a broad postmortem likeness right is dangerous. A rule that survives their objections has to be narrow — about consent and disclosure, not ownership and takedowns. A harder bill to write. Also a better one.
What this actually means for you
Not policy. You, this week, at your own kitchen table.
Write down what you want, while you can. The researchers' term is premortem intent; the plain version is a paragraph in your will, or a letter to whoever handles your estate, saying whether you consent to a synthetic version of your voice, face or writing. "No, never" is a valid answer and takes one sentence.
Ask your parents. Out loud. Once. Ten minutes, awkward for about ninety seconds. Far cheaper than guessing later.
Read the terms on the products already holding the material — photo library, voice recordings, messaging archive, any legacy-contact feature. Look past the privacy blurb to the transfer clause: what happens to this data in an acquisition, and who can end the arrangement.
Keep your own copy. Whatever recordings actually matter, hold them somewhere you control, not only inside a service. That takes a company out of the middle of your memory.
If you contact a legislator, ask the narrow question. Not "ban deepfakes" — that fight is stuck. Ask: does an ordinary person get to decide, in advance, whether they may be simulated after death, and who enforces it? California decided for performers. Nobody decided for your father.
The lesson, as I see it
The griefbot debate keeps getting framed as a question about grief — whether it is healthy, whether the mourners are doing it wrong. That framing is a mistake, and a convenient one, because it puts bereaved people on trial instead of a business model. The real question is not whether talking to a simulation of your father is good for you; reasonable people differ, the evidence is thin, and frankly it is neither my business nor yours. The real question is whether he agreed — and right now, in most of the world, nobody knows, because nobody asked, because there is no mechanism to ask.
That is fixable, and it isn't even hard. Denmark drafted a version and its answer sits in a drawer. California wrote one for people with agents. Neither is finished — but both prove this can be written down. The obstacle was never technical.
My vote? Not ownership — consent. Ask people while they can answer, record the answer somewhere it survives them, and make the default silence mean no rather than help yourself. Everything else — duration, enforcement, who inherits what — is detail worth arguing over for a decade. The asking is not.
I still have the voicemail. I am not deleting it, and I am not building anything out of it either. It gets to stay what it is: a man running late, worrying about my parking, forty seconds long, and finished.
Ask the people you love what they'd want done with their voice — clumsily, over dinner, before it becomes a legal question. Then send them this, so they know you meant it. The HAIA Foundation works to get these questions asked while somebody can still answer; the rest is on our Substack.