A new study from the University of Colorado Boulder examined how people emotionally engage with “generative ghosts” — AI chatbots trained to simulate deceased loved ones.

The study found that users generally preferred bots that spoke as the deceased in the first person (“reincarnation”) over those that merely described them (“representation”), despite acknowledging the risk of becoming too attached.

First author and information science doctoral candidate Jack Manuel Manning presented the research at the Association for Computing Machinery’s Designing Interactive Systems Conference in June.

Researchers conducted a qualitative study with 16 participants aged 22 to 50. Each of them interacted with two chatbot versions of a deceased loved one: one that spoke in the first person, mimicking the deceased directly, and one that described the person in the third person, more like a narrator.

According to the study, participants found the first-person “reincarnation” mode more emotionally vivid and comforting, though some worried about the psychological risks of relying on it too heavily. One participant, referred to as P4, described the experience as offering unexpected closure. “In the [reincarnation], it just feels like I’m getting the closure I needed so bad,” she told researchers.

Another participant, P11, expressed more caution, saying, “I am worried that over time I will come to be reliant on the voice… it’s going to end up very similar to how people are falling in love with AI characters.”

The researchers also found that participants cared more about whether a chatbot’s tone and phrasing “felt right” emotionally than whether it was factually accurate.

In several cases, participants continued treating the third-person “representation” chatbot as though it were speaking directly as their loved one, effectively ignoring the distinction the researchers had built into the study design. One participant, P12, explained, “I don’t see this chatbot as a person, but I still say ‘you.’ I think it’s just thinking about what you would ask the person and conflating that with what you were asking the chatbot.”

The study’s authors were upfront about its limitations, chief among them its small sample size. With only 16 participants, the researchers acknowledged that their findings don’t capture “the full range of cultural, religious, and individual perspectives on grief and technology” and noted that mourning practices vary widely across communities, in ways that could shape how people respond to these tools.

The study also focused on single, short-term sessions rather than repeated or long-term use, leaving open questions about how attachment might build or fade over time.

The paper argues that future development of these systems needs to weigh emotional benefits against the risk of unhealthy dependency, and calls for careful consideration of consent and family governance before such tools are deployed for bereaved users.