Designing Conversations with the Dead: How People Engage with Generative Ghosts

Problem
This preprint addresses the gap in understanding user interactions with generative AI systems designed to simulate deceased individuals, termed “generative ghosts.” The literature lacks comprehensive qualitative insights into how different conversational frameworks—specifically, third-person representation versus first-person reincarnation—affect user experience, authenticity, and emotional engagement. The study aims to elucidate these dynamics, which are critical for the design of AI systems that engage with sensitive topics such as death and memory.

Method
The authors conducted a qualitative user study involving 16 participants who interacted with generative ghosts under two distinct conversational frameworks: representation (third-person) and reincarnation (first-person). The study employed semi-structured interviews to gather data on participants’ experiences, focusing on dimensions such as authenticity, affect, and perceived risks. The analysis was grounded in thematic coding, allowing for the identification of key patterns in user preferences and emotional responses. The study emphasizes the collaborative nature of interactions, highlighting how users’ memories and emotional connections to the deceased influence their engagement with the AI.

Results
Participants exhibited a clear preference for the reincarnation mode, citing its immediacy and emotional resonance. However, they also expressed concerns about potential over-reliance on the AI for emotional support. In contrast, the representation mode was favored for its ability to facilitate engagement with memories rather than direct conversational presence, although many participants blurred the lines between the two modes, often engaging in dialogue despite the third-person framing. The study found that affective resonance consistently outweighed factual fidelity in user interactions, indicating that emotional connection is paramount in these contexts. The findings suggest that conversational tone, language, and rhythm—shaped by users’ memories—significantly influence the quality of interactions with generative ghosts.

Limitations
The authors acknowledge several limitations, including the small sample size of 16 participants, which may not capture the full diversity of user experiences. The study is also limited by its qualitative nature, which may restrict generalizability. Additionally, the focus on a specific demographic may introduce biases that affect the findings. The authors do not address the potential ethical implications of creating generative ghosts, such as the psychological impact on users or the societal consequences of AI representations of deceased individuals.

Why it matters
This research has significant implications for the design of AI systems that engage with sensitive topics, particularly in the realms of grief and memory. Understanding user preferences for conversational frameworks can inform the development of more empathetic and effective generative models. The findings highlight the importance of emotional engagement over factual accuracy, suggesting that future AI systems should prioritize affective design elements. This work opens avenues for further exploration into the ethical considerations and psychological impacts of AI interactions with representations of the deceased, which are increasingly relevant in a digital age where such technologies are becoming more prevalent.

Authors: Jack Manning, Daniel Sullivan, Dylan Thomas Doyle, Anthony T. Pinter, Jed R. Brubaker
Source: arXiv:2605.21390
URL: https://arxiv.org/abs/2605.21390v1