Information scientist Jack Manning and colleagues reported study results on June 12 in the proceedings of the Designing Interactive Systems Conference. The study involved 16 volunteers who conversed with AI-generated simulations of deceased loved ones.

Volunteers reported that conversations with the AI simulations were often comforting, even when the chatbot invented facts. Manning is an information scientist at the University of Colorado Boulder.

"They love this technology," Manning said. Participants entered the study with specific intentions for their interactions. Study volunteers reported having a specific goal, such as an unanswered question or a detail they wanted to share, prior to the interaction.

"They had a question they felt was unanswered or a detail in their life they were desperate to share," he said. Volunteers reported the experience was positive across the board. "It felt like I was actually talking with my grandpa," a study participant said.

"It just feels like I’m getting the closure I needed so bad," a study participant said. The chatbot used only a brief survey completed by the participant about their deceased loved one to generate impersonations. Participants often continued interactions when the bot fabricated details, such as mentioning a job the deceased never held.

Participants were more bothered by mistakes in tone or style than by factual errors. The chatbot did not use emojis, which mattered to participants who had primarily communicated with their loved ones via text. In one instance, a bot addressed a participant as "champ," a term the participant stated their loved one never used.

People who did not want to interact with a re-creation of a deceased loved one were excluded from the study. A human researcher mediated each conversation in the study by reviewing the chatbot’s replies before passing them to participants. The human mediation served as a guardrail to prevent interactions that might be deceptive or harmful.

One study participant expressed concern about becoming too reliant on conversations with the bot. "I don’t know if I would like the person I would become if I kept using this," a study participant said. He expressed concern regarding the lack of consistent guardrails for AI ghost technology outside of the study environment.

Manning’s sister died when he was 11 years old. He stated he has a strong aversion to the idea of talking with a digital simulation of his sister. "I have a version of my sister in my head," he said.

"I wouldn’t want to be confronted with the fact that maybe I’m wrong in my way of remembering her," he said. The study results reduced his personal aversion to AI-generated ghosts. He stated that carefully designed generative ghosts could be a positive tool for some people working through grief.

"There is a version of generative ghosts that could be a really positive thing in the world," he said. Tomasz Hollanek is an AI ethicist at the University of Cambridge. Hollanek was not involved in the research. Companies currently sell services that recreate deceased loved ones through AI-generated text, voices, or faces.

The study used a locally hosted language model to generate responses, ensuring participant data was not sent to external servers. Each simulation was active for a single session and then permanently deleted. Participants ranged in age from 24 to 68, with relationships to the deceased including parents, grandparents, and siblings. Some participants reported feeling emotional distress during the interaction, though all said they would participate again. The research team emphasized that the simulations were not intended to replace human grief counseling. No participant reported believing the simulation was actually their deceased relative. The team plans to conduct follow-up studies with larger and more diverse groups. Ethical guidelines for such technologies remain under development, with experts calling for regulation to prevent exploitation of vulnerable users.