Griefbots: The Ethics of Talking to an AI Version of the Dead
AI recreations of the dead can now be called, texted and video-chatted. The harder questions are about grief that never quite closes, and a person who never got to say yes.
Most of us keep one somewhere: a voicemail we can't delete, a voice note, a text thread that stops mid-conversation. A recording has one hard, gentle property. It says what it says and nothing more. You can replay it a hundred times, and on the hundred-and-first it will still be the same sentences in the same voice. That sameness is part of the ache, and part of the comfort too, because it lets the loss stay a fact.
Griefbots remove that property. Feed a system someone's old voice memos, texts and photos and it will answer you, in something like their voice, with sentences they never said. Platforms now offer calls, texts and video chats with these recreations, and 2026 reporting describes a digital-afterlife industry that is growing fast. I don't think the right response is horror or enthusiasm. It's a set of questions, and most of them are worth asking before you press the button rather than after.
What a Griefbot Actually Is
The idea isn't new. One of the best-known companion apps began as a founder's private project, after a close friend died, to rebuild the friend from old text messages so the conversation could go on. What's new is the polish: voices that sound right, faces that move, and responses fluent enough to feel like a person paying attention.
Writing software is my job, so let me say the unglamorous part. What comes out of one of these systems isn't the person. It's plausible. Plausible is the cheapest thing a language model produces, and it will happily invent a sentence your father would never have said and deliver it warmly, in his cadence, with no sign that anything was made up. The system doesn't know him. It knows patterns in what he left behind, and it completes them.
That matters for grief in a specific way. When you replay a voicemail, you know exactly where the boundary is. When a griefbot answers, the line between what they actually said and what a model produced starts to blur, and memory, which is already a reconstructive process, may absorb the difference.
Grief and the Line That Stays Open
Here I want to be fair to the technology, because modern grief research doesn't say the goal is to let go and move on. Since the 1990s, an influential view has been "continuing bonds": many bereaved people keep an inner relationship with the person who died, talking to them, wondering what they'd think, marking their birthdays, and this is normal and often healthy. People have always written letters to the dead. In that light, a griefbot could look like a letter that writes back.
The difference is the writing back. A letter, a photograph, a grave, a voicemail: none of them ever produces anything new. Part of what grief involves is coming to terms with exactly that. Researchers describe a back-and-forth most mourners go through, turning toward the loss and then taking a break from it, until over time the loss becomes something you can carry. A griefbot can supply an endless stream of new material from someone who has died, and that keeps the relationship technically open. "Ambiguous loss" is a term the family therapist Pauline Boss coined for situations like a missing person or dementia, where someone is both present and absent. It wasn't written with this in mind, but it points at the same unease: a loss with no clear edge is hard to finish.
The honest state of the evidence is that there isn't much. The technology is newer than the studies. There's no settled finding that griefbots help or harm, and I'd distrust anyone who claims one. What can be said is that some uses might be useful, such as a way to say what went unsaid, or company through a hard anniversary. And a few risks are worth naming: avoiding the loss instead of moving through it, a bot that says something painful or false, and the second loss that arrives when a company shuts down, changes its pricing, or updates the model so the person you knew changes overnight.
Who Said Yes?
The consent problem is the one I find hardest to set aside. The person being simulated never agreed to it, and can't. Researchers who study the digital-afterlife industry have flagged this, and a 2024 paper from Cambridge researchers argued for consent protocols covering both the people being simulated and the people who'd interact with the bots, along with ways to retire a simulation with dignity.
Think about what a person's voice contains. How your mother spoke to you isn't how she spoke to a stranger, or to her doctor, or to herself. A model built from scattered fragments flattens all of that into one plausible voice, and it may end up saying things in her name that she'd have found mortifying. The question isn't only privacy after death. It's dignity: whether she'd recognize herself, and whether she'd have agreed to be put in front of her grandchildren as an approximation.
It gets murkier when a simulation is built from scraped public data such as social posts, interviews and videos, perhaps by someone who wasn't close to the person at all. Consent then isn't merely missing, it was never asked for. And it's rarely one person's decision. A widow may want the bot while a daughter doesn't. The training data can include messages from other people who wrote to the deceased, and they didn't agree to be in the model either. I haven't verified the current state of the law in any particular country, and it's changing, but I'd be surprised if it had settled these questions.
Healthy Use and Prolonged Use
Clinicians have a name for grief that doesn't ease with time: prolonged grief disorder, added to the DSM-5-TR in 2022. For adults it involves intense yearning or preoccupation with the person who died, lasting at least 12 months after the loss and disrupting daily life. It's a diagnosis for a clinician, not a self-diagnosis, and it isn't my intent to suggest that anyone who talks to a griefbot has it. There's no evidence that a griefbot causes it. The worry is narrower: that an always-available simulation could make it harder to reach the point where the loss becomes bearable, particularly for people already at risk.
The framing that comes up when bereavement researchers and death doulas discuss this, and I'm describing a general view rather than quoting anyone, is that the difference lies in the shape of the use, not in the technology. Healthy use tends to have edges: a purpose, a time, a way to stop, and other people and things in the rest of the day. A ritual has a beginning and an end. Prolonged use looks like the reverse. The days get organized around the conversation. Turning it off feels unthinkable. Time with the people who also knew the person shrinks, because they're less available and less agreeable than the bot. The question that view keeps returning to is a plain one: what is this for?
Questions to Ask Before You Begin
These are for anyone considering a griefbot for someone they've lost, and, in the second list, for the practical version of the same question: what terms you'd want set for your own data.
If you're considering one for someone who died
- What am I hoping it will do? Hear their voice again, say something unsaid, avoid the silence in the house? Each has a different answer, and the unsaid thing may be better written in a letter that doesn't answer.
- Would they have wanted this? Did they ever say anything about it? If you don't know, what's the cost of getting it wrong?
- Who else has a stake? Family members, and the people whose messages would end up in the data.
- What will it be built from, and who holds it? Where do the recordings go? What happens if the company closes or changes its terms?
- How will I stop? Decide the limit before you start, whether that's a date, a number of conversations or a purpose, and tell someone you trust.
- What will I do if it says something they never would have said? Eventually it will. Decide beforehand how you'll treat that moment.
If you're setting terms for your own data
- Write down whether you'd want to be simulated at all, by whom, for what, and for how long. "I'd rather be remembered than replicated" is a perfectly good instruction.
- Name who has the authority to say yes or no, and who can switch it off.
- Tell them, and keep it with your will or digital estate notes, the way you'd say who gets the photographs.
- Check the legacy settings on the accounts that hold your voice and messages. Most large platforms have a legacy contact or an inactive-account option. It's dull, and it takes ten minutes.
FAQ
Is a griefbot the same as an AI companion? They're related, but a companion is an invented character. A griefbot imitates a specific real person who can no longer speak for themselves, and that's where the extra questions about consent and memory come from.
Is it wrong to use one? Not necessarily. Keeping a bond with someone who died is normal, and some people may find a bot helpful. The research is early, so it makes sense to go in with a purpose, a limit and an honest look at whose permission is missing, especially if your grief already feels complicated.
Can a griefbot cause prolonged grief disorder? There's no evidence that it does. The concern is that an always-available simulation might make it harder to accept that the person won't say anything new. That's a reason for caution, not a finding.
What if the person left recordings on purpose? Then the consent question changes, since they took part. It's still worth asking whether they imagined being listened to or being talked with, because those are different things.
What if I've already used one and feel stuck? Talk to a bereavement counselor or a hospice grief-support program, many of which offer free support. You haven't done anything shameful. Some silences are hard to sit in, and wanting them to answer back is about as human as grief gets.