What the Study Found
- Participants preferred first-person “reincarnation” AI ghosts over third-person “representation” modes, citing greater immediacy and emotional closeness.
- Users judged authenticity by emotional tone, dialect, and conversational rhythm — not factual accuracy; a wrong term of endearment could break the illusion.
- Even when the AI framed replies in the third person, participants often responded as if addressing the deceased loved one directly.
- All 16 participants said they’d use the technology again, but many feared grieving loved ones could grow emotionally dependent on it.
On a Zoom screen in a Boulder, Colorado, laboratory, a 32-year-old woman is typing to her grandmother, who’s been dead for five years. Her grandmother is typing back. “I can see her. I can feel her,” the woman writes to the researchers. “It just feels like I’m getting the closure I needed.”
Off camera, Jack Manning is watching this happen and recalibrating. He’d come in braced for revulsion; instead he was getting closure.
Manning, a PhD candidate at the University of Colorado Boulder, and his supervisor Jed Brubaker have published what they believe is the first controlled user study of “generative ghosts”, AI chatbots trained on data about the dead so that the living can chat with them. The technology is not speculative any more. Companies such as Project December, Séance AI and HereAfter AI already sell it, stitching journal entries, texts and voice recordings into interactive versions of your lost people; a handful now offer full virtual-reality holograms you can walk beside. Almost nobody, until this month, had bothered to ask real grievers what it actually feels like to use one.
The lab that lets you text the dead
The setup was low-tech. Sixteen participants, aged 22 to 50, logged in for a short interview with a facilitator who took a kind of oral history: what the deceased sounded like, how they held themselves, which endearments they used, which they never would. In the background, a second researcher (cameras off, playing operator) fed those details into GPT-4 and steered its replies. Then the participant would type into the chat window and, roughly a minute later, a message would come back. From grandpa. From mum. From the friend who died two years ago.
Everyone spoke to two versions of the ghost, whose behaviour is documented this month in the Proceedings of the 2026 Designing Interactive Systems Conference. One narrated the deceased in the third person, a sort of digital family archivist. The other spoke as the deceased in the first person. The team called the first mode representation and the second reincarnation, braced themselves for a backlash against the more aggressive version, and didn’t get one. Across the board, participants praised reincarnation for its immediacy, its intimacy, its sense of an old conversation resuming after a long pause. A 50-year-old woman, whose grandmother had once promised in life to visit her after death, seemed to feel the promise being kept. “It was so so powerful,” she typed. “I’d like for you to come to me again.”
“We originally thought it might feel very Black Mirror creepy to people and make them uncomfortable,” Manning says. “I ended up being completely wrong. People thought it was amazing.”
Tone beats accuracy, every time
What stood out from the transcripts is how loosely participants held on to fact. The hallucinations that plague any large language model were waved through, forgiven, ignored. What couldn’t be forgiven was tone. When one participant’s stepfather-ghost greeted him with “champ”, a word the real man had never used, the illusion buckled instantly and the participant nearly walked out of the session; another was pulled up short when her grandmother-ghost began sprinkling in Spanish phrases her grandmother never would have used. The right endearment mattered more than any right anecdote. Affective accuracy, as the researchers call it, trumped factual accuracy nearly every time.
Rhythm mattered, too. AI has a way of producing dense, thoughtful paragraphs, and dense, thoughtful paragraphs are almost nobody’s idea of a text from grandma. Users wanted short lines, emojis, the comfortable back-and-forth of a phone conversation with someone who loves you. Anything longer felt like distance.
There were stranger slippages. Faced with the third-person representation ghost, several participants ignored the framing altogether and started addressing it directly, as if the biographer weren’t in the room; nobody did the reverse. One participant even found representation the more truthful of the two, because her grandmother in life had tended to talk at her rather than with her, and the distant, monologuing bot, by accident, was an excellent likeness.
Then came a twist. Asked whether they would use the technology again, every single participant said yes. Asked what they worried about, they mostly worried about someone else: a brother who might get addicted trying to extract answers the dead can’t really give, a friend who might use a ghost as a substitute for anyone still alive. The theme was not that this was dangerous. It was that it might be dangerous for someone more fragile than the person saying so.
Manning knows the pull. He lost his sister to a heart condition when they were both kids, and spent years looking for a richer way to keep her present. When he first heard about griefbots he was horrified, which, he says, is exactly why he thought he should study them. The alternative was to leave an intensely intimate technology to the people most eager to sell it.
“I felt it was important for me to do the work because the people who are the largest fans might skip the empirical research and just make a product,” Manning says. He adds, of the technology’s potential to harm: “I think a lot about 11-year-old me. If I had access to ChatGPT and it started responding as my sister late at night without supervision…that is a very scary thought.” Then, of its potential to help: “But as we have learned through this paper, it can also be an incredibly meaningful experience for people that enables them to get some closure and peace.”
Brubaker’s lab is already deep into follow-up, including a study with clinical mental-health professionals to work out what a responsible griefbot should and shouldn’t do; the commercial products, of course, are not waiting. Voice clones, hologram walks, subscription models: the ghost economy is here already, and what this study suggests is that the instinct most of us reach for, that talking to a simulacrum of the dead is grotesque, doesn’t really survive contact with the bereaved.
The people in the lab did not want reverence. They wanted a specific word for love and a reply short enough to feel like a text back, and whether it is wise to give them that, on a Tuesday night, alone, for as many hours as they can bear, is a question the market is preparing to answer whether we are ready or not.
- Study type: Qualitative user study, within-subjects design with AI-assisted Wizard-of-Oz setup
- Intervention: Chat-based interactions with LLM-generated “generative ghosts” of a deceased loved one, seeded with participant-provided descriptions
- Comparator: Reincarnation mode (AI speaks as the deceased in first person) vs. representation mode (AI speaks about the deceased in third person)
- Sample size: 16 participants, ages 22–50, each of whom had lost a close relative or friend
- Duration: Single ~45–90 minute session; approximately 20 minutes per ghost mode
- Funding / conflicts of interest: Supported by the National Science Foundation and a Google Academic Research Grant; no conflicts of interest disclosed
- Peer-review status: Peer-reviewed; published in the Proceedings of the 2026 Designing Interactive Systems Conference (DIS ’26; 24% acceptance rate)
- Main limitation: Authors note the small, culturally narrow sample and the short, one-off encounter design, which cannot speak to how attachment or dependency might evolve with repeated long-term use
Reference
Manning, J. M., Sullivan, D., Doyle, D. T., Pinter, A. T., & Brubaker, J. R. (2026). Designing Conversations with the Dead: How People Engage with Generative Ghosts. In Proceedings of the 2026 Designing Interactive Systems Conference (pp. 3610–3621). ACM. DIS ’26: Designing Interactive Systems Conference. https://doi.org/10.1145/3800645.3813090
Frequently Asked Questions
Why did people in the study prefer AI ghosts that spoke in the first person?
People preferred AI ghosts that spoke in the first person because that voice created a sense of immediacy, as though an old conversation had simply resumed rather than been described from a distance. Participants said it felt less like reading a biography of their loved one and more like actually being with them, which was where the emotional payoff of the experience seemed to live.
Is it true that talking to a “griefbot” of a dead relative is inherently disturbing?
Talking to a griefbot of a dead relative is not inherently disturbing, at least not to the people who tried it. In the Colorado Boulder study, the researchers expected participants to find the experience creepy but nearly all of them described it as comforting, and every single one said they would use the technology again. What unsettled participants was not the idea in the abstract; it was the thought of someone more vulnerable than themselves using it without limit.
How does an AI griefbot actually convince someone it sounds like their dead loved one?
An AI griefbot convinces someone it sounds like their dead loved one mostly through tone, rhythm and word choice, not through accurate biographical detail. The Boulder participants forgave the model when it hallucinated facts, but the illusion collapsed the instant it used a wrong endearment or a phrase the real person never would have used. Short, texting-style replies with the right cadence felt authentic; long, essay-length answers felt like distance.
Could using an AI ghost of the dead interfere with grieving?
Using an AI ghost of the dead could interfere with grieving, and the participants themselves raised this concern, though almost always on behalf of someone else. They worried that a more fragile person, such as a still-grieving sibling, might become dependent on the ghost or use it as a substitute for living relationships. The researchers are now running follow-up studies with clinical mental-health professionals to work out what a responsible griefbot should and should not do.
What’s stopping AI ghosts of the dead from going mainstream?
Very little is stopping AI ghosts of the dead from going mainstream; commercial products from companies such as Project December, Séance AI and HereAfter AI are already on the market, and some now offer voice clones and virtual-reality holograms. The open question is not technical but ethical, whether these systems can be designed to comfort the bereaved without fostering unhealthy attachment, and whether the market will wait for that answer before scaling further.
Cite This Page