Meta holds a United States patent, granted December 30, 2025, covering the simulation of a user from their posts and voice. A patent shows intent, not a shipped product.
The products that already exist go by two names—griefbot and deadbot—and they do the same thing. Each one stands in for someone who died. An AI chat or voice avatar trains on that person’s messages, recordings, and whatever else they left behind in digital form. The patent facts above were checked at one remove, from the patent record rather than from Meta.
The commercial market for these is small, and it shifts fast enough that any list of companies dates quickly. Several earlier entrants are gone. Project December sells text sessions, while HereAfter AI and StoryFile record living users for later playback after they die.
Those two models aren’t the same product. One builds a griefbot from material the person never handed over, and one plays back a recording they sat down to make.
Services divide into consent-by-design and post-hoc creation, and the divide runs deeper than the technology. Post-hoc creation needs no donor at all. Anybody holding a pile of old messages can build a griefbot of the person who wrote them, and no donor-consent mechanism exists. Unwanted contact from a griefbot the family never asked for is a worry that researchers and users both raise.
The archive that would train a griefbot is the same archive an estate saves after a death. Both sit in the same accounts. Pulling those copies down is covered on save the photos and files.
The evidence runs to one small interview study
The evidence runs thin in both directions. One small peer-reviewed interview study of real users exists, and it’s the whole located base for claims about them.
That study interviewed 10 participants, five of whom had talked with griefbots of their own dead. The sample skews three ways. Its participants were mostly in the United States, the sample skewed male, and everybody volunteered. Findings were mostly positive, self-reported, and described as time-limited use during acute grief.
Clinical outcomes are a different matter, and this study measured none of them at all. No located study has. Standard grief instruments, control groups, and follow-up periods are all absent from the work on real griefbot users.
So the honest summary is short. The study talked to 10 people, five of them had the experience, and nobody has yet measured what it does to grief.
That gap matters, because confident claims about these services circulate in both directions. Neither direction has data. A service offering comfort has no measured outcome behind it, and a warning about harm has none behind it either. This site takes no position on whether anybody should use one.
What the study did find is worth stating with care, because care is the thing this topic tends to lack. Use was time-limited and transitional. People described turning to a griefbot during acute grief and then moving on from it, in their own self-reported words.
A study that size can’t carry more weight than that. A family deciding about this in 2026 is deciding without data, and no amount of vendor confidence changes that for them.
Four concerns come up, and none has been tested
Researchers and the people they interviewed raise four concerns, and none of them has been tested against outcomes. Dependency heads the list. Disrupted grief processing comes next—the ordinary work of grieving interrupted rather than carried forward. Getting the person’s identity wrong follows, where a griefbot misses the real person in ways a family plainly notices.
Unwanted contact closes the list, where a griefbot reaches somebody who never agreed to meet one. None of the four has been measured. Reporting them as documented harms would repeat the mistake this page is trying hardest to avoid.
These are concerns rather than findings, and the difference matters. Saying so doesn’t dismiss them, because untested worries are the normal state of any genuinely new product.
The recommendations run the other way. Cambridge AI ethicists Hollanek and Nowaczyk-Basińska call for donor consent before death and adult-only access. They also want a documented route for retiring a griefbot. None of that binds anybody, because no provider has to follow a recommendation from a research group.
A 2025 perception study found the same wish among ordinary people rather than among professional ethicists. Presented at the CHI conference, it reported that people want advance control over data scope and over who gets access. They also wanted a trusted person to manage their material after death, and nobody gets appointed to that by default.
Both papers point at the same missing piece. The wish is documented and the mechanism isn’t, and that gap is where a family planning ahead ends up standing.
No federal law governs these services
Rules here are largely absent, and no federal statute in the United States governs these services. One state law comes close. California’s Assembly Bill 1836, passed in 2024 and effective January 1, 2025, creates liability for unconsented digital replicas of a deceased personality. The amount is the greater of $10,000 or actual damages.
Two textual limits sit inside that, and both come from the reading of the analysts who supported the bill. No case law has turned up. A commercial-value threshold likely puts ordinary private people outside it, and the statute reaches expressive audiovisual works and sound recordings.
Chat-based services sit awkwardly against that wording. Whether the statute reaches a text-only griefbot is textually doubtful, and that’s a question for a lawyer rather than for a website.
Assembly Bill 1836 is a right-of-publicity statute, which protects the commercial value of a name, a voice, and a likeness. It isn’t a general consent law. Nothing in it gives a family a broad say over whether somebody builds a griefbot of a relative. The distance between what the statute covers and what a family might expect from it is wide.
The estate-planning fact behind all of this is short enough to state in a single line. A documented advance decision is what exists. As of July 28, 2026, no United States law makes even a recorded refusal binding on a provider.
That doesn’t make a refusal worthless. It gives a family something to point at and a provider something to weigh, which is more than silence gives.
Where you write the refusal down matters, because a will filed for probate becomes a public court record. A letter of instruction stays private. Estate documents grant authority, and the private letter carries the detail about recordings, messages, and what you want done with them. What belongs in a will sets out that division.
The advance decision is the part a person controls
The archive decides the question long before any service does, because a griefbot needs raw material to train on. Who holds the messages holds the option. The settings that decide who receives a copy are the same ones deciding the rest of a digital estate.
Three settings outrank a will. Facebook’s legacy contact, Google’s Inactive Account Manager, and Apple’s Legacy Contact outrank it under RUFADAA, the model state law on digital assets.
Each one sets its own scope, and the scopes differ more than you’d expect. Facebook’s contact can’t read messages. Google’s Inactive Account Manager sends a data-download link to as many as 10 recipients after a set period of silence. Apple’s legacy account leaves out purchased media and the iCloud Keychain, so the stored passwords stay behind.
Message content is the material a griefbot would need, and it’s the material these tools guard most closely. That guarding is deliberate. The three settings that outrank your will covers what each one does.
A recorded preference has no legal force over a provider. It still tells the person managing your estate what you wanted, which is the part a family can act on.
Naming somebody to manage this after death is what the CHI participants said they wanted. Nobody gets appointed by default. The person who ends up holding the message archive holds the choice, whether or not anybody meant them to. So the choice is worth making on purpose, rather than by accident of who had the password.
Write the preference down somewhere, tell the person who will hold the archive, and say plainly what you want done. Expect no provider to be bound by it. As of July 28, 2026, that document is what a family has to work with, and it’s more than most families are left with.