A griefbot is what it sounds like: an AI chatbot or voice assistant built to mimic someone who has died, trained on whatever digital trace they left behind — text messages, emails, photos, voice notes, social media posts. Some are text-only. Some, using voice cloning, can hold a spoken conversation in the dead person’s actual voice. A few, paired with video generation, can produce something closer to a video call.
The technology has moved from speculative to commercially available faster than the ethical and legal frameworks around it. And the raw material these services need — photos, voice recordings, written communication — is exactly what most people store in their cloud photo and file accounts without a second thought about what it could someday be used to build.
How a Griefbot Actually Gets Made
The process is more straightforward than the emotional weight of the result suggests. A family member or friend supplies a griefbot service with whatever material is available: text message and email exports, social media history, photos with captions, and — increasingly central to making the result feel convincing — voice recordings, even short ones, which voice cloning technology can use to generate new speech in that person’s voice, saying things they never actually said.
The bot is then built on a large language model fine-tuned or prompted to adopt the deceased person’s speech patterns, personality traits, and conversational style, inferred from the supplied material. The person interacting with it experiences something designed to feel like continuity — a way to keep talking to someone who’s gone.
The Consent Problem That Can’t Be Patched
Every other use of someone’s personal data carries at least the theoretical possibility of consent: a company can ask you to agree to terms, you can decline, you can later revoke that agreement. Griefbots break this model in a way that has no real precedent.
The person being recreated cannot consent, object, or correct the record. They didn’t agree to have an AI trained on their voice and personality after death. If the bot’s output misrepresents them — saying something they’d never have said, adopting a tone that doesn’t match who they were — they have no way to say “that’s not me.” Bioethicists studying this technology have flagged this specific asymmetry as the central ethical problem: every other subject of AI training data retains some agency over how they’re represented. A deceased person trained into a griefbot has none.
Most people never imagined their photos and voice notes would be used this way. When someone records a voice note for a friend, or has their photo taken at a birthday dinner, they’re not consenting to that material becoming training data for a future chatbot built after their death. There’s no terms-of-service checkbox for this scenario, because it didn’t exist when most of the source material was created.
The data is intensely intimate — and not just for the deceased. Text conversations, voice notes, and photos used to train a griefbot almost always include other people: the friends and family who were part of those conversations and photos. Their words, their voices, their likenesses get pulled into the training material too, without them having been asked either.
Who Gets to Decide?
This is where the legal and family disputes start, and they’re not hypothetical. Multiple family members may have different views on whether a griefbot of a deceased relative should exist at all — one sibling wanting it as a source of comfort, another finding the idea of an AI imitation of their parent disturbing or disrespectful. Estate law in most jurisdictions wasn’t written with this scenario in mind: a person’s photos and voice recordings are property that passes to heirs, but whether those heirs can or should use that property to build a conversational simulation of the person is largely unaddressed by existing wills, terms of service, or platform policies.
Some platforms have started updating their terms of service to address posthumous use of account data, but coverage is inconsistent, and most people have never read the specific clause that would determine whether their own photos and voice notes could end up training a griefbot after they’re gone.
The Exploitation Risk for the Grieving
Beyond consent, there’s a second concern that bioethicists and consumer advocates raise consistently: grief is a state of heightened emotional vulnerability, and griefbot services are, in most cases, commercial products. Some charge ongoing subscription fees to keep “talking” to a recreated loved one. The combination of genuine grief, the appeal of continued connection, and a recurring charge creates conditions where exploitation is a real risk, not just a theoretical one — people in acute grief are not always in the best position to evaluate whether a paid service is helping them process loss or keeping them suspended in it.
Why Voice Notes Specifically Changed the Calculus
Text-based recreations of deceased people have existed in some form for years — chatbots trained on someone’s writing style aren’t new. What changed the ethical and emotional stakes substantially is the maturity of voice cloning technology, which now requires only a short sample — sometimes under a minute of clear audio — to generate new speech convincingly mimicking someone’s actual voice.
This matters specifically for personal storage habits because voice notes are exactly the kind of content people increasingly save without much thought: a quick message to a partner, a recorded birthday greeting, a voicemail saved because deleting it felt wrong. None of these were ever treated as significant data by the person recording them. Under the griefbot framework, every one of them is potential training material — short, casual, and unguarded in a way that makes the resulting clone feel unsettlingly authentic to people who hear it.
The emotional intensity of hearing a cloned voice say new words is, by most accounts from people who’ve experienced griefbot services, qualitatively different from reading text generated in someone’s writing style. Voice carries identity in a way text doesn’t fully capture — which is exactly why it’s also the component of this technology that demands the most scrutiny.
What Regulation Has and Hasn’t Caught Up To
A few jurisdictions have begun addressing digital likeness rights after death, generally as an extension of existing “right of publicity” laws that historically covered commercial use of a deceased celebrity’s image or voice. But these laws were written with commercial endorsement in mind — stopping a company from using a dead celebrity’s face to sell a product without estate permission — not the more intimate, non-commercial scenario of a family member building a private griefbot of a parent or spouse.
The result is a regulatory gap that mirrors the early years of most new technology categories: existing law addresses the adjacent, more obvious case (commercial exploitation of a famous person’s likeness) while leaving the actual emerging use case (private, personal recreation of an ordinary person using personal data) largely unaddressed. Until that gap closes, the practical protection for most people’s photos and voice notes comes down to the policies of whoever is storing that material — not a law written for a different problem.
What This Means for How You Store Your Own Memories
You don’t need to have an opinion on whether griefbots should exist to draw a practical conclusion from this: the photos and voice notes you store today are the exact raw material this entire industry runs on, and where you store them — and what the storage provider’s policies actually allow — determines whether they could end up there without anyone in your family explicitly deciding that’s what should happen.
A few things worth checking, regardless of which storage provider you use:
Does the provider’s terms of service grant rights to use your content for purposes beyond storage? Broad licenses to “improve services” or “develop new features” using your content can, depending on how they’re written, extend further than most people assume when they accept them without reading.
Is there a clear policy for what happens to your account after death? Some services have legacy contact or memorialization features; others default to indefinite account dormancy or eventual deletion after inactivity, with no clear path for family decision-making either way.
Does the provider train any AI — its own or a third party’s — on your content? This is the question that matters most for the griefbot scenario specifically. A provider with a clear, structural commitment not to use your content for AI training of any kind is one where your voice notes and photos can’t quietly become training material for anything, during your life or after.
Where daftei Stands
daftei never trains AI models — its own or any third party’s — on content stored in your vault. That commitment doesn’t change based on whether the account is active, dormant, or has been inherited; it’s a structural constraint on the business, not a feature toggle. Your photos and voice notes are stored to be retrieved by you, encrypted in transit with TLS 1.3 and at rest with AES-256, and that’s the entirety of what they’re used for.
Account deletion follows a 30-day grace window before permanent, irreversible erasure — a policy that applies the same way whether you initiate it yourself or whether it’s eventually initiated by someone managing your account after you’re gone. We don’t currently offer a dedicated legacy-contact feature, and if that matters to you, it’s worth having an explicit conversation with the people who’d be managing your accounts about what you want to happen to your stored content.
A Question Worth Sitting With
Griefbots aren’t going away — the underlying technology is now accessible enough that the question isn’t whether this category of product will exist, but how it gets governed, and whose photos and voices end up training it without anyone having explicitly agreed to that outcome.
You can’t fully control what happens to your digital footprint after you’re gone. But you can be deliberate about where your most personal material — the photos, the voice notes, the recordings of your actual voice saying actual things — lives while you’re here, and whether the provider holding it has made training AI on it impossible by policy, or merely unlikely by current practice. Those are very different guarantees.