Generative AI systems that impersonate deceased loved ones can provide genuine comfort to mourners, even when the chatbots fabricate biographical details about the people they are meant to represent, a finding that surprised the researchers themselves.
The study, presented at the 2026 Designing Interactive Systems (DIS) Conference and published in its proceedings, asked 16 volunteers to converse with AI-generated “ghosts” of people they had lost. Each participant completed a brief survey about their loved one that covered personality, interests, and shared memories, and the researchers built a simple chatbot from that limited input.
The result: every participant reported the experience as positive.
“It felt like I was actually talking with my grandpa,” one participant said.
Another: “It just feels like I’m getting the closure I needed so bad.”
Accuracy is not the point
The most striking finding may be what the participants did not care about. The chatbots, working from minimal data, routinely fabricated biographical details. They invented jobs, hobbies, and anecdotes without hesitation. Participants barely noticed, and when they did, they did not mind.
“They love this technology,” said Jack Manning of the University of Colorado Boulder, the study’s lead author, who described himself as “very surprised” by the result.
What did bother participants was tonal mismatch. One bot called a participant “champ”, a nickname the participant said the loved one would never have used. The chatbot also never used emojis, a significant detail for participants who had texted frequently with the deceased. Errors in tone derailed the experience in ways that factual errors did not.
Manning offered an explanation: “I have a version of my sister in my head. I wouldn’t want to be confronted with the fact that maybe I’m wrong in my way of remembering her.”
The implication is that AI ghosts succeed not by replicating objective truth about a person, but by meeting emotional and social expectations. The technology’s role is less about accurate simulation than about creating a space for continued relationship, a digital extension of the internal conversations many people already have with their dead.
Guardrails in the study
The experiment was conducted under carefully controlled conditions. Every chatbot reply was reviewed by a human researcher before being shown to the participant. This ensured that even if the bot produced something potentially harmful, such as a raw reminder of death or a promise that could not be kept, a real person intercepted it.
The commercial landscape is far less regulated. Anyone can prompt a chatbot to impersonate a specific person. Companies already sell AI recreations with synthesized voices and animated faces. There are no consistent guardrails across these products.
The participants themselves recognized the risks. One wondered whether they might become “too reliant” on the bot. Another said: “I don’t know if I would like the person I would become if I kept using this.”
An ethical path forward
Tomasz Hollanek, an AI ethicist at the University of Cambridge who was not involved in the study, said the research provides a valuable starting point: “If we want to understand how to design these systems ethically, we need more evidence of this kind.”
The key ethical tension is straightforward. Grief technology that genuinely helps people is not inherently wrong. Manning himself described his view as softened by the results, seeing “a version of generative ghosts that could be a really positive thing in the world.” But the same technology, deployed without guardrails, could deepen grief, distort memory, or create unhealthy dependencies.
The study suggests that careful design, human mediation, transparency about the system’s limitations, and attention to tone over factual accuracy, could tip the balance toward benefit.
For now, the research opens a window onto a question that will only grow more urgent as AI companions become more convincing: if a conversation with a dead person cannot be real but feels real, is that comfort worth embracing, or a risk to guard against?
Reference: Manning, J. et al. (2026). “Designing conversations with the dead: How people engage with generative ghosts.” Proceedings of the 2026 Designing Interactive Systems Conference. DOI: 10.1145/3800645.3813090
News article: Hulick, K. (2026). “AI ‘ghosts’ can comfort mourners, even when the bots get the facts wrong.” Science News. Available at: sciencenews.org