I can see her. I can feel her. It just feels like closure.
In a laboratory at CU Boulder, researchers are doing what grief has always quietly asked of us — finding ways to keep the dead present among the living. Using artificial intelligence trained on the digital traces of those who have passed, scientists Jack Manning and Jed Brubaker are studying how sixteen bereaved participants respond to conversations with generative ghosts, discovering that people seek not mere representation but something closer to resurrection. The technology already exists commercially, and the researchers are racing to understand its psychological terrain before it becomes as ordinary as a phone call.
- A grieving woman types to her dead grandmother and feels closure; a man nearly walks out when the AI calls him by the wrong nickname — the emotional stakes of this technology are immediate and deeply personal.
- Commercial platforms are already selling access to the dead, training AI on journals, texts, and voice recordings before researchers have mapped the psychological consequences.
- All sixteen study participants unanimously preferred AI ghosts that spoke as the deceased in first person, revealing a human hunger not for memory but for the illusion of continued presence.
- Researchers warn that vulnerable, grief-stricken individuals could develop unhealthy dependencies on these digital ghosts, and new studies with mental health professionals are now underway.
- The technology is moving faster than the science meant to govern it, and the window to build ethical safeguards is narrowing with every product launch.
On the second floor of a CU Boulder building, researchers Jack Manning and Jed Brubaker have created a space for conversations with the dead. A grieving volunteer speaks with a facilitator while a second researcher feeds details into an AI system, training it in real time to speak like the deceased. Within minutes, the participant is exchanging messages with a ghost.
The results, published in the Proceedings of the 2026 Designing Interactive Systems Conference, were not what the researchers expected. Rather than finding the experience unsettling, all sixteen participants — ages twenty-two to fifty — found it profound. When given the choice between an AI that spoke as the deceased in first person and one that described them from the outside, every participant chose resurrection over representation. They wanted the ghost to say "I remember" rather than "she remembered."
The details proved critical. Participants forgave small factual errors but recoiled at wrong terms of endearment — one man nearly ended his session when the AI called him "champ," a word his stepfather would never have used. Emotional tone, sentence rhythm, even emoji use mattered enormously. The closer the ghost sounded to the actual person, the more meaningful the exchange felt.
The technology is already commercial. Platforms like Project December and HereAfterAI train AI on social media, journals, voice recordings, and photographs, allowing the bereaved to interact with multimedia versions of the dead. Some companies offer fully immersive virtual-reality experiences. Brubaker believes generative ghosts will soon be ordinary — which is precisely why he believes they must be studied now.
Manning came to this research through personal loss: his sister died of a heart condition when they were children. His initial horror at AI ghosts convinced him he was the right person to study them rigorously, rather than leave the field to enthusiasts who might skip the empirical work. He thinks often about his eleven-year-old self, alone at night with access to an AI trained as his sister — a prospect that still frightens him, even as the research has shown him the technology's genuine capacity for comfort.
Every participant said they would use the technology again, but nearly all added the same caveat: they worried about vulnerable people accessing these ghosts without guidance. The lab has already begun new studies with mental health professionals, trying to chart both the benefits and the dangers before the dead become, as the researchers put it, a permanent part of how the living go on.
On the second floor of a building at CU Boulder, Jack Manning and Jed Brubaker have built a laboratory for conversations with the dead. The setup is straightforward: a volunteer sits down for a video call with a facilitator who asks questions about someone they've lost. While the conversation happens, another researcher feeds those details into an artificial intelligence system, training it in real time to speak like the deceased. Within minutes, the grieving person is typing messages to a ghost.
One woman, thirty-two years old, found herself in text-based conversation with her grandmother, dead five years. "I can see her. I can feel her," she wrote. "It just feels like I'm getting the closure I needed." Another, fifty, typed back to the memory of her beloved: "It was so so powerful. I'd like for you to come to me again." These exchanges, documented in a paper published this month in the Proceedings of the 2026 Designing Interactive Systems Conference, represent the first rigorous scientific look at what researchers call generative ghosts—AI agents trained on the digital traces of people who have died.
What Manning and Brubaker discovered surprised them. They had expected the technology to feel unsettling, something out of a dystopian film. Instead, participants found it profound. The study involved sixteen people, ages twenty-two to fifty, each of whom had lost someone close. Each person interacted with two versions of an AI ghost: one that spoke in the first person, as if the dead person had returned, and another that spoke in the third person, describing the deceased from outside. The preference was unanimous. People wanted resurrection, not representation. They wanted the ghost to say "I remember going to the beach together," not "She loved going to the beach with you."
The details mattered enormously. Participants were forgiving of occasional errors—the small hallucinations that language models sometimes produce. But they recoiled if the ghost used the wrong term of endearment. One man nearly ended his session when the AI version of his stepfather called him "champ," a word the real man would never have used. Users also preferred shorter sentences punctuated with emojis to the long, rambling paragraphs that AI systems naturally generate. Emotional tone, dialect, conversational rhythm—the more the ghost sounded like the actual person, the more meaningful the interaction felt.
The technology is no longer theoretical. Platforms like Project December and Séance AI train text-based ghosts on journal entries, social media posts, and text messages from the dead. HereAfterAI goes further, inviting people to record their own voices and submit photographs so that after they die, their loved ones can interact with a multimedia version of them. Some companies have even built fully immersive virtual-reality experiences, allowing the bereaved to walk alongside a hologram of someone they've lost. Brubaker, an associate professor of Information Science, believes generative ghosts will soon become ordinary. But he also believes they need to be studied carefully before they do.
Manning came to this work through personal loss. His sister died of a heart condition when they were children, and he had spent years searching for meaningful ways to hold her memory. When he first heard about AI ghosts, he was horrified—which is precisely why he knew he was the right person to study them. "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," he said. He thinks often about his eleven-year-old self, alone late at night with access to an AI trained to respond as his sister. That prospect frightens him. But the research has also shown him something else: the technology can offer genuine closure and peace.
At the end of the study, researchers asked participants whether they would use the technology again. Everyone said yes. But almost all of them added a caveat: they worried what would happen if vulnerable people—those still raw with grief, those prone to unhealthy attachment—got access to these ghosts without guidance. The lab has already begun new studies, working with mental health professionals to map both the benefits and the dangers. Manning and Brubaker are trying to understand the technology before it becomes commonplace, to build safeguards before the need for them becomes urgent. The dead, it seems, are about to become very much a part of how we live.
Notable Quotes
We originally thought it might feel very Black Mirror creepy to people and make them uncomfortable. I ended up being completely wrong. People thought it was amazing.— Jack Manning, PhD candidate in Information Science at CU Boulder
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. But it can also be an incredibly meaningful experience for people that enables them to get some closure and peace.— Jack Manning
The Hearth Conversation Another angle on the story
Why did people prefer the first-person version? Wouldn't it feel more honest to talk about someone rather than pretend they're still here?
That's the intuitive objection, but the study showed something different. When the ghost spoke as "I," it felt like actual presence—like the person had come back. When it spoke as "she" or "he," it felt like talking to a memorial, a description. People wanted the former.
Even though they knew it wasn't real?
Yes. They knew it was AI. But the emotional experience of being addressed directly, of having the ghost remember things in first person, created something the third-person version couldn't match. It was the difference between a eulogy and a conversation.
What about the "champ" moment? That seems like it could have gone either way.
It did. The man nearly walked away. But it also revealed something crucial: people have very specific memories of how someone spoke to them. Get that wrong, and the whole thing collapses. The ghost has to sound like the actual person, not a generic version of them.
So this is really about fidelity to the individual?
Exactly. The more accurate the emotional tone, the dialect, the rhythm of speech—the more people felt they were actually connecting with someone they'd lost, not just talking to a chatbot.
And they all wanted to do it again?
All of them. But almost all of them also said they were worried about other people using it. They could see the potential for harm—for someone to become dependent on talking to a ghost instead of grieving in healthier ways.
So the researchers are caught between two truths?
Yes. The technology can offer real closure and peace. But it can also become a way to avoid grief entirely. That's why more research is needed before this becomes mainstream.