The Grief That Doesn't Move

Grief psychologists have known for decades that mourning does not mean severance. The dominant clinical model since the 1990s, "continuing bonds," replaced an older Freudian idea that healthy grief requires cutting emotional ties to the dead entirely. In practice, people keep talking to lost parents in the car, dress in a late spouse's sweater, ask a dead grandmother what she'd think of a decision. One bioethicist studying grief technology puts the baseline at about 96 percent of mourners: nearly everyone already maintains some relationship with the deceased through memory, objects, dreams, or one-sided conversation. None of that required AI. It's what grieving has always looked like from the inside.

What AI griefbots — also called deadbots, or more gently "generative ghosts" — do is take that one-sided conversation and answer it back, in a voice trained on the dead person's texts, emails, and recorded speech. And the research arriving this year is more interesting than either the comforting or the alarming version of this story that tends to circulate on its own.

What the studies actually found

A 2026 study out of the University of Colorado Boulder, published as "Designing Conversations with the Dead," built two versions of a chatbot for sixteen grieving volunteers: a "reincarnation" bot that spoke in the deceased's first person, and a "representation" bot that spoke about them in the third person, more like a well-informed narrator than a resurrection. Participants supplied a description of the person — relationship, personality, how they talked — and had roughly twenty-minute conversations with each. The emotional register split cleanly along that design choice: the first-person bot produced immediacy and closeness, sometimes startling closeness, while the third-person version produced something closer to reflection and memory-work. One volunteer described the experience as "getting the closure I needed so bad." Tellingly, factual accuracy mattered less to participants than whether the bot's language actually sounded like the person — a reminder that what mourners are evaluating isn't correctness, it's voice.

A separate qualitative study, conducted with 25 bereaved adults in China and published in Frontiers in Digital Health, found the same split held under interview, not just survey. Participants described real psychosocial benefit — comfort, a sense of unfinished business addressed, continued connection — sitting right next to real unease: a feeling of inauthenticity, worry about dependency, and concern that the technology could disrupt the natural course of grieving rather than support it. The researchers' summary framing is worth sitting with: griefbots were "neither inherently beneficial nor inherently harmful," because the comfort and the risk come from the exact same mechanism — the simulated presence of someone who is gone.

The mechanism, named

That mechanism has a name in the newer literature, and it's the specific thing that should give a technology reporter pause: several groups, including bioethicists at the University of Cambridge's Leverhulme Centre for the Future of Intelligence, warn that deadbots can fuel complicated or prolonged grief — a recognized clinical condition in which mourning doesn't taper off on its own timeline but gets stuck. Continuing bonds theory was always premised on those bonds evolving: the relationship you carry with a dead parent at year one is not the relationship you carry at year ten, and that slow change is part of how grief resolves. A chatbot trained to reproduce someone's voice as it was at the moment of death has no mechanism for evolving with you. It can only get better at sounding exactly like who they used to be — which is a strange kind of stasis to install inside a process whose whole logic is movement.

The Cambridge researchers, writing in Philosophy & Technology, raise a second problem that has nothing to do with grief psychology and everything to do with consent: a deadbot can be built and given to someone as a gift, with neither the deceased's permission while alive nor the recipient's consent to receive it. They describe the realistic risk of "unwanted digital hauntings" — a returning presence someone didn't ask for and may not be able to simply turn off. Their proposed safeguards read like the design brief nobody in the griefbot industry has been required to write: opt-out by default, a ban on repurposing a deadbot for advertising or an ongoing social media presence, and — a detail that says more than it seems to — a "dignified retirement" mechanism, something like a digital funeral, so the bot can be deliberately ended rather than just left running indefinitely because ending it feels like a second death.

What this adds up to

None of the new research argues that griefbots are a bad idea in the way, say, unmoderated companion chatbots for isolated teenagers might be. The comfort participants describe is real, consistently reported, and consistent with what grief psychology already expected from continuing-bonds behavior generally. The risk is also real, and it's oddly specific: not that talking to a simulation of the dead is inherently unhealthy, but that a simulation with no design for its own ending asks a mourner to do something continuing bonds was never built to do — maintain a relationship that doesn't change, with someone who no longer can.

That is the same shape this site keeps finding in domain after domain: the tool doesn't invent a new human need, it takes a real one — connection, comfort, unfinished conversation — and removes the friction that used to regulate it. With writing, the friction was struggle; with reading, it was the pause before the answer; with grief, it may simply be time. Letters fade, photographs go quiet, and even memory reorganizes itself without being asked. A griefbot doesn't have to. Whether that turns out to be a mercy or a trap looks, so far, less like a fixed property of the technology and more like a design decision nobody has been forced to make carefully yet.

Sources

1. Designing Conversations with the Dead: How People Engage with Generative Ghosts — arXiv / University of Colorado Boulder (Preprint)
2. Exploring public perceptions of artificial intelligence in bereavement support: a qualitative study of griefbots — Frontiers in Digital Health (Peer-Reviewed)
3. Deadbots and the Moral Knowledge of Grief — Philosophy & Technology, University of Cambridge Leverhulme Centre for the Future of Intelligence (Peer-Reviewed)
4. Move Over Chatbot, Here Comes the Deadbot — CORDIS, European Commission (Research Communication)
5. AI 'deadbots' can fuel pathological grief and affect how we deal with death — TechXplore (Journalism)
6. Griefbots Are Here, Raising Questions of Privacy and Well-being — The Hastings Center (Bioethics Commentary)