If you could chat with an AI ghost, what would you want them to say? New study explores
Inside a second-floor lab in CU 91Ҹ’s Information Science building, Jack Manning and Jed Brubaker are quietly resurrecting the dead.
As the researchers look on, study volunteers log into Zoom and chat at length with AI-generated representations of lost grandparents, siblings, parents and family friends.
Some get emotional: “I can see her. I can feel her,” said one 32-year-old woman during a text-based conversation with her grandmother who died five years ago. “It just feels like I’m getting the closure I needed.”
Others plan the next visit: “It was so so powerful,” typed a 50-year-old woman to the ghost of her beloved. “I’d like for you to come to me again.”
These interactions, chronicled in a paper published this month by the , offer the first scientific glimpse at how people use “generative ghosts”—the increasingly popular AI agents trained on data about the deceased.
Among other things, the study found that participants preferred ghosts that spoke in the first person, acting as a resurrection rather than a representative. The more accurate and life-like the ghost’s emotional tone, dialect and conversational rhythm, the better.
“We originally thought it might feel very Black Mirror creepy to people and make them uncomfortable,” said first author Manning, a doctoral candidate in information science who found his way to the unusual field of study after losing his sister. “I ended up being completely wrong. People thought it was amazing.”
Jack Manning, a doctoral candidate in the Department of Information Science, presents his research findings at the ACM Designing Interactive Systems conference in Singapore. Credit: Jack Manning
Conjuring the dead
Once viewed as science fiction, generative ghosts, sometimes called “griefbots” or “deathbots,” are fast becoming a commercial reality.
Platforms like Project December and use journal entries, social media posts and texts from the deceased to train text-based ghosts for surviving loved ones to chat with. Companies like invite users to submit voice recordings and photos of themselves to create multimedia ghosts for their loved ones to interact with after they’re gone. Some start-ups have even , enabling grieving clients to literally walk with a hologram of the dead.
Brubaker, an associate professor of information science who has spent most of his career at the intersection of tech and death, it won’t be long before generative ghosts are a regular part of life as we know it.
But given their potential to both help and harm the grieving, he believes they should be designed with solid research as a guide.
That’s where his lab comes in.
“To our knowledge, we are conducting the first user experience studies of simulated AI ghosts,” said Brubaker.
Jang Ji-Sun interacts with an AI simulation of her late daughter Na Yeon in a documentary by a South Korean broadcaster. Credit: MBC Media/YouTube
Don't call me 'champ'
For their inaugural study, the research team recruited 16 people, ages 22–50, who had lost a close relative or friend.
Participants logged into Zoom for a brief on-camera interview with a facilitator, who gathered biographical information and other details about the deceased. Meanwhile, in the background, a second researcher plugged this information into a large language model (LLM), building a ghost in real-time.
Participants chatted with two iterations, each for about 20 minutes. One ghost spoke in first-person (“I remember going to the beach together.”) Another used third person. (“She loved going to the beach with you.”)
The facilitator stood by to intervene if things got uncomfortable, while the operator behind the scenes used the conversation details to fine-tune the ghost.
Subsequent interviews found that, across the board, participants preferred what the researchers called the “reincarnation” over the “representation,” even pivoting to address the latter as if they were their loved one.
For instance, if a ghost said, “Is there something you want your grandpa’s advice on?” the participant would respond: “Grandpa, what would you want to tell me if you were here?”
While people seemed willing to overlook occasional inaccuracies, or "hallucinations" spun up by AI, they cringed if the bot used the wrong term of endearment. (When the ghost of one participant’s stepfather called him "champ"—a term he would never have used—the participant nearly called off the session.)
“Maybe the chatbot didn’t know when exactly grandma was born, but if it used some slang or regional word choice that she used, it was extremely impactful,” said Brubaker.
Users also preferred shorter sentences with emojis rather than the rambling paragraphs that AI tends to produce.
In search of closure
Information Science Professor Jed Brubaker
Perhaps the most interesting finding came at the end, when participants were asked if they would use the technology again.
Surprisingly, everyone said yes (although some feared they would become overly reliant on it). But almost all added that they feared what would happen to their grieving loved ones if they got their hands on one.
“I just fear he would get addicted,” said one study subject about a brother still struggling with a loss.
Manning came to the study with skepticism.
He had lost his sister to a heart condition when they were kids and had longed for years for more meaningful ways to memorialize her.
“There were bake sales and concerts, all to try to create a place where we could come together to remember her, but it just seemed like a monumental task,” he said.
When he heard about AI ghosts, he was initially horrified—which is exactly why he felt he was a good fit 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,” Manning said.
The lab has already begun the next studies, including one working with mental health professionals to analyze the benefits and risks of interacting with AI ghosts.
Along with their potential pitfalls, Manning now sees their promise, too.
“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,” he said. “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.”