Grief is one of the most universal human experiences, and it has always resisted easy solutions. Now, a growing category of commercial technology is positioning itself as a form of comfort: AI-powered chatbots designed to simulate conversations with deceased loved ones. A study published on June 12 in the Proceedings of the 2026 Designing Interactive Systems Conference offers some of the first structured empirical evidence about what actually happens when bereaved people interact with these systems. The findings are nuanced, and that nuance is exactly what most coverage of this topic tends to flatten.
Two Bots, Two Very Different Emotional Experiences
The study recruited 16 volunteers who had experienced a loss and were willing to engage with an AI-generated, text-based re-creation of the deceased. Before each session, participants provided researchers with a description of the person they had lost, covering personality traits, communication style, and the nature of their relationship. That information was then used to configure what researchers call “griefbots” or “deadbots,” built on an AI model that previously powered ChatGPT.
Each participant interacted with two distinct versions of the bot for approximately 20 minutes each, via Zoom. The first version, called the reincarnation bot, spoke in the first person, producing messages like “I remember when we went to the beach together.” The second, called the representation model, used the third person: “She loved going to the beach with you.”
This design choice turned out to matter enormously. The reincarnation model tended to generate feelings of immediacy and closeness. The representation model, by contrast, prompted memory and reflection. These are not the same emotional states, and they are not interchangeable. The distinction suggests that how an AI ghost speaks is not a stylistic detail but a design decision with real psychological consequences.
What Worked, What Didn’t, and Why Authenticity Mattered More Than Accuracy
Here is something that might seem counterintuitive: the bots sometimes fabricated facts about the deceased, and participants largely did not care. Inaccuracies about specific details did not significantly undermine the overall experience. What did matter, sometimes intensely, was linguistic authenticity. Participants reacted negatively when the bot used terms of endearment that the deceased person would never have used, words like “champ” or “sweetie.” Conversely, when the bot captured a regional expression or a characteristic phrase, the effect could be striking.
Jed Brubaker, an information scientist at the University of Colorado Boulder and a co-author of the study, described this dynamic directly: a chatbot that did not know a grandmother’s exact birth date but reproduced her distinctive word choices could still feel deeply real. This is a meaningful finding. It suggests that emotional fidelity operates on different terms than factual fidelity. People are not looking for a database of biographical facts. They are looking for a recognizable presence.
Overall, participants responded positively to the interactions. One volunteer who had lost her mother described the experience as “getting the closure I needed so bad,” adding that she could feel her mother’s presence through the conversation. Co-author Jack Manning, also at CU Boulder, noted that participants arrived with specific goals: unanswered questions, things they had wanted to say, details they were desperate to share.
The Ethical Terrain That Research Is Only Beginning to Map
The emotional benefits documented in the study exist alongside concerns that participants themselves raised. Some worried about developing emotional dependency on the bots. Others expressed discomfort at the idea of interactions that might feel socially or ethically out of place. One participant articulated the risk plainly: prolonged use might change who they become as a person.
Researchers share these concerns. Nora Lindemann, a technology ethicist at Osnabrück University in Germany, has pointed to what she describes as the crucial danger: that people who use these tools may not need to adjust to a world without the deceased person, remaining instead in what she calls a “somewhat pretend, in-between stage.” This is not a fringe concern. It sits at the center of what makes grief bots genuinely difficult to evaluate.
Commercial products in this space already exist. Companies including Project December, Re;memory, and Séance AI are operating in this market, even as the psychological and ethical questions remain largely unresolved. Manning, who lost his own sister at age 11 and was initially skeptical of the technology, framed his motivation for conducting the research in terms of responsibility: without empirical study, he argued, those most enthusiastic about the technology might simply build products without examining their effects. Tomasz Hollanek, a technology ethicist at the University of Cambridge who was not involved in the study, echoed this position, calling for more evidence to guide ethical design.
In Short
AI grief bots are not a hypothetical future technology. They are a present-day commercial reality, and people are already using them. This study is one of the first to examine what that experience actually involves. The key takeaways: the way a bot speaks (first person versus third person) shapes the emotional experience in distinct ways; linguistic authenticity matters more than factual accuracy; and participants, while largely positive, are already aware of the risks. The technology can provide genuine comfort. It can also, if used without care, complicate the very process it claims to support. That tension is not going away, and research like this is how society begins to navigate it responsibly.
Based on reporting from Smithsonian Magazine.