pith. sign in

arxiv: 2605.21390 · v1 · pith:BN4TXSP2new · submitted 2026-05-20 · 💻 cs.HC · cs.AI

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

Jack Manning , Daniel Sullivan , Dylan Thomas Doyle , Anthony T. Pinter , Jed R. Brubaker This is my paper

Pith reviewed 2026-05-21 03:24 UTC · model grok-4.3

classification 💻 cs.HC cs.AI
keywords generative ghostsdigital afterlifeAI representation of deceasedreincarnationaffective resonancecollaborative designqualitative user study
0
0 comments X

The pith

People prioritize emotional resonance over factual accuracy when conversing with AI versions of the deceased.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper investigates two design choices for generative ghosts, AI systems trained on data from deceased individuals: representation, where the AI discusses the person in the third person, and reincarnation, where it speaks in the first person as if it were them. In a qualitative study with 16 participants, the authors find that users often disregard the distinction between these modes and value interactions that evoke personal emotional connections based on their memories. They conclude that such engagements are inherently collaborative, influenced by unique factors like tone and conversational rhythm rather than strict factual correctness.

Core claim

Participants in the study engaged with generative ghosts by privileging affective resonance over factual fidelity across both representation and reincarnation modes. Reincarnation was favored for its sense of immediacy, yet raised concerns about over-reliance, while representation allowed focus on memory but was frequently treated as direct conversation anyway. The interactions are shaped by elements unique to the user's personal recollection of the deceased, making them always collaborative in nature.

What carries the argument

The representation versus reincarnation distinction in generative ghost design, which the study shows users often blur in favor of emotionally resonant, collaborative dialogues.

If this is right

  • Reincarnation offers immediacy but carries risks of over-reliance on the AI.
  • Representation facilitates engagement with memory, yet users frequently engage in direct dialogue anyway.
  • Tone, language, and conversational rhythm unique to the user's memory of the deceased shape the interactions.
  • All interactions with generative ghosts are collaborative.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Designers of such systems should emphasize capturing personal emotional and stylistic elements over comprehensive biographical data.
  • This collaborative aspect might extend to other AI personifications, suggesting users co-create the experience.
  • Potential for therapeutic use but requires careful consideration of emotional risks like grief dependency.

Load-bearing premise

The study assumes that self-reported experiences from a small group of 16 participants in a qualitative setting accurately reflect broader user preferences and that the representation versus reincarnation distinction remains meaningful even when participants frequently disregard it in practice.

What would settle it

A quantitative study with more participants demonstrating that factual fidelity is prioritized over affective resonance in interactions with generative ghosts would falsify the main claim.

read the original abstract

We examine how people experience two choices in the design of generative ghosts, AI systems that are trained on data of the dead: representation, where an AI speaks about a deceased person in the third person, and reincarnation, where the AI speaks as the deceased in the first person. Through a qualitative user study with 16 participants, we explore how each shaped authenticity, affect, and risk. Reincarnation was preferred for its immediacy, but participants shared fears of over-reliance. Representation was preferred for engaging with memory over conversational presence, though participants often ignored this distinction, engaging in dialogue despite third-person framing. Across both modes, participants privileged affective resonance over factual fidelity. We conclude by showing how factors such as tone, language, and conversational rhythm -- factors unique to the user's memory of the deceased -- shape interactions with generative ghosts, and argue that those interactions are always collaborative.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript reports on a qualitative user study with 16 participants examining two design modes for generative ghosts (AI systems trained on data of the deceased): representation, in which the AI speaks about the deceased in the third person, and reincarnation, in which the AI speaks as the deceased in the first person. The study explores effects on authenticity, affect, and risk, reporting that reincarnation was preferred for immediacy (with fears of over-reliance), representation for engaging memory (though participants often ignored the distinction), and that affective resonance was privileged over factual fidelity across modes. The authors conclude that interactions are always collaborative and shaped by user-specific factors such as tone, language, and conversational rhythm.

Significance. If the thematic findings hold, the work offers timely HCI insights into affective and collaborative dimensions of AI-mediated engagement with the deceased, potentially informing design choices around personal memory elements. The qualitative approach surfaces plausible participant concerns around over-reliance and the blurring of representation versus reincarnation. However, the small sample and absence of quantitative validation or saturation evidence constrain the strength of the preference and universal-collaboration claims.

major comments (2)
  1. Abstract and concluding section: the assertion that 'those interactions are always collaborative' and that participants 'privileged affective resonance over factual fidelity' is a load-bearing generalization drawn from self-reported themes in a sample of 16 participants who frequently disregarded the representation/reincarnation framing. The manuscript provides no details on theoretical saturation, negative-case analysis, or variation across demographics that would support extending these observations beyond the specific cohort and interface.
  2. Methods and findings sections: the preference claims rest on qualitative self-reports without quantitative measures, larger-scale validation, or checks against interface/prompt artifacts, weakening support for the stronger conclusions about user preferences and risk perceptions.
minor comments (1)
  1. Abstract: the limitations of the n=16 sample and the noted participant disregard for the core distinction could be stated more explicitly to better contextualize the findings.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. We address each major comment below, clarifying the scope of our qualitative findings while revising claims and methods descriptions where appropriate to better reflect the study's exploratory nature and limitations.

read point-by-point responses

  1. Referee: Abstract and concluding section: the assertion that 'those interactions are always collaborative' and that participants 'privileged affective resonance over factual fidelity' is a load-bearing generalization drawn from self-reported themes in a sample of 16 participants who frequently disregarded the representation/reincarnation framing. The manuscript provides no details on theoretical saturation, negative-case analysis, or variation across demographics that would support extending these observations beyond the specific cohort and interface.

    Authors: We agree that the current phrasing risks overgeneralization. We will revise the abstract and conclusion to qualify these observations explicitly as emerging from our participant cohort (e.g., 'participants in our study privileged affective resonance...' and 'our findings indicate that such interactions are collaborative in nature, shaped by...'). We will also expand the methods section to describe our reflexive thematic analysis process in greater detail, including how themes were iteratively developed, attention to negative cases, and the absence of formal saturation testing given the study's exploratory aims. Demographic variation will be noted as a limitation rather than claimed as representative. revision: partial

  2. Referee: Methods and findings sections: the preference claims rest on qualitative self-reports without quantitative measures, larger-scale validation, or checks against interface/prompt artifacts, weakening support for the stronger conclusions about user preferences and risk perceptions.

    Authors: Our study was designed as an in-depth qualitative exploration suited to surfacing nuanced affective and authenticity experiences in an emerging HCI domain; quantitative validation would require a separate, larger-scale experiment. We will add explicit discussion in the methods of steps taken to mitigate prompt and interface artifacts (e.g., consistent base prompts and post-interview debriefs) and will strengthen the limitations section to acknowledge that self-reported preferences are not triangulated with behavioral metrics. This does not alter our view that the thematic insights remain valuable for design implications. revision: partial

Circularity Check

0 steps flagged

No circularity: qualitative claims derived directly from participant data without reduction to inputs or self-citations

full rationale

The paper reports a qualitative study with 16 participants examining two design modes for generative ghosts (representation vs. reincarnation). Core conclusions—that participants privileged affective resonance over factual fidelity and that interactions are always collaborative—are presented as outcomes of thematic analysis of interview data on authenticity, affect, and risk. No equations, fitted parameters, predictions, or first-principles derivations appear in the provided text. The abstract and description contain no self-citations invoked as load-bearing uniqueness theorems or ansatzes that would reduce the central argument to prior author work by construction. The study is self-contained against its own empirical inputs; any concerns about sample size or generalization fall under interpretive scope rather than circular derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The paper rests on standard assumptions of qualitative HCI research and introduces the term generative ghosts without independent external validation of the concept.

axioms (1)
  • domain assumption Self-reported preferences from a small number of participants in a study setting can reveal meaningful patterns about authenticity, affect, and risk in AI interactions.
    This underpins the extraction of themes on preference and over-reliance from the 16 interviews.
invented entities (1)
  • generative ghosts no independent evidence
    purpose: Label for AI systems trained on data of the dead that enable conversational interaction.
    New framing introduced to organize the two design modes under study.

pith-pipeline@v0.9.0 · 5689 in / 1311 out tokens · 48166 ms · 2026-05-21T03:24:17.572538+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages

  1. [1]

    A. A. Adams and S. A. Williams. 2014. What’s yours is mine and what’s mine’s my own: joint accounts and digital identity.ACM SIGCAS Computers and Society44, 1 (April 2014), 15–26. doi:10.1145/2602147.2602150

    work page doi:10.1145/2602147.2602150 2014

  2. [2]

    Gerd Andersson, Kristina Höök, Dário Mourão, Ana Paiva, and Marco Costa. 2002. Using a Wizard of Oz study to inform the design of SenToy. InProceedings of the 4th Conference on Designing Interactive Systems: Processes, Practices, Methods, and Techniques(London, England)(DIS ’02). Association for Computing Machinery, New York, NY, USA, 349–355. doi:10.1145...

    work page doi:10.1145/778712.778762 2002

  3. [3]

    2017.Death and digital media

    Michael Arnold, Martin Gibbs, Tamara Kohn, James Meese, and Bjorn Nansen. 2017.Death and digital media. Routledge. Manuscript submitted to ACM 18 Manning et al

    work page 2017

  4. [4]

    Eric PS Baumer and Jed R Brubaker. 2017. Post-userism. InProceedings of the 2017 CHI Conference on Human Factors in Computing Systems. 6291–6303

    work page 2017

  5. [5]

    Timothy Bickmore and Justine Cassell. 2001. Relational agents: a model and implementation of building user trust. InProceedings of the SIGCHI Conference on Human Factors in Computing Systems(Seattle, Washington, USA)(CHI ’01). Association for Computing Machinery, New York, NY, USA, 396–403. doi:10.1145/365024.365304

    work page doi:10.1145/365024.365304 2001

  6. [6]

    Patrick Biernacki and Dan Waldorf. 1981. Snowball Sampling: Problems and Techniques of Chain Referral Sampling.Sociological Methods & Research 10 (1981), 141–163. doi:10.1177/004912418101000205

    work page doi:10.1177/004912418101000205 1981

  7. [7]

    2012.Thematic analysis.American Psychological Association

    Virginia Braun and Victoria Clarke. 2012.Thematic analysis.American Psychological Association

    work page 2012

  8. [8]

    2018.Protocol for Responding to Participant Distress (Adapted Version for Telephone

    Jed R Brubaker. 2018.Protocol for Responding to Participant Distress (Adapted Version for Telephone. Technical Report. Video Communication. Technical Report. University of Colorado Boulder

    work page 2018

  9. [9]

    Brubaker and Vanessa Callison-Burch

    Jed R. Brubaker and Vanessa Callison-Burch. 2016. Legacy Contact: Designing and Implementing Post-mortem Stewardship at Facebook. InProceedings of the 2016 CHI Conference on Human Factors in Computing Systems. ACM, San Jose California USA, 2908–2919. doi:10.1145/2858036.2858254

    work page doi:10.1145/2858036.2858254 2016

  10. [10]

    Brubaker, Lynn S

    Jed R. Brubaker, Lynn S. Dombrowski, Anita M. Gilbert, Nafiri Kusumakaulika, and Gillian R. Hayes. 2014. Stewarding a legacy: responsibilities and relationships in the management of post-mortem data. InProceedings of the SIGCHI Conference on Human Factors in Computing Systems. ACM, Toronto Ontario Canada, 4157–4166. doi:10.1145/2556288.2557059

    work page doi:10.1145/2556288.2557059 2014

  11. [11]

    We will never forget you [online]

    Jed R Brubaker and Gillian R Hayes. 2011. " We will never forget you [online]" an empirical investigation of post-mortem myspace comments. In Proceedings of the ACM 2011 conference on Computer supported cooperative work. 123–132

    work page 2011

  12. [12]

    Jed R Brubaker, Gillian R Hayes, and Paul Dourish. 2013. Beyond the grave: Facebook as a site for the expansion of death and mourning.The Information Society29, 3 (2013), 152–163

    work page 2013

  13. [13]

    Jed R Brubaker, Meredith Ringel Morris, Dylan Thomas Doyle, Casey Fiesler, Martin Gibbs, and Joanna McGrenere. 2024. AI and the Afterlife. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems. 1–5

    work page 2024

  14. [14]

    Juliet Corbin and Anselm Strauss. 2008. Basics of Qualitative Research (3rd ed.): Techniques and Procedures for Developing Grounded Theory. (2008). doi:10.4135/9781452230153

    work page doi:10.4135/9781452230153 2008

  15. [15]

    Dylan Thomas Doyle and Jed R Brubaker. 2023. Digital legacy: a systematic literature review.Proceedings of the ACM on Human-Computer Interaction 7, CSCW2 (2023), 1–26

    work page 2023

  16. [16]

    I Am So Overwhelmed I Don’t Know Where to Begin!

    Dylan Thomas Doyle and Jed R Brubaker. 2024. " I Am So Overwhelmed I Don’t Know Where to Begin!" Towards Developing Relationship-Based and Values-Based End-of-Life Data Planning Approaches. InProceedings of the CHI Conference on Human Factors in Computing Systems. 1–14

    work page 2024

  17. [17]

    Dylan Thomas Doyle, Casey Paul, and Jed R Brubaker. 2025. Assessing Support for Mortality: An Environmental Scan of Online Platforms. Proceedings of the ACM on Human-Computer Interaction9, 2 (2025), 1–18

    work page 2025

  18. [18]

    Jason Fagone. 2021. He couldn’t get over his fiancee’s death. so he brought her back as an A.I. Chatbot. https://www.sfchronicle.com/projects/2021/ jessica-simulation-artificial-intelligence/

    work page 2021

  19. [19]

    Xue Fan and Xiantong Cheng. 2025. Adoption of AI-Based Deathbots: Behavioral Reasoning Theory Perspective on Drivers, Barriers, and Ethical Attitude.International Journal of Human–Computer Interaction(2025), 1–24. doi:10.1080/10447318.2025.2543995

    work page doi:10.1080/10447318.2025.2543995 2025

  20. [20]

    Robert Douglas Ferguson, Michael Massimi, Emily Anne Crist, and Karyn Anne Moffatt. 2014. Craving, creating, and constructing comfort: insights and opportunities for technology in hospice. InProceedings of the 17th ACM Conference on Computer Supported Cooperative Work & Social Computing (Baltimore, Maryland, USA)(CSCW ’14). Association for Computing Machi...

    work page doi:10.1145/2531602.2531631 2014

  21. [21]

    Kun Fu, Chenxi Ye, Zeyu Wang, Miaohui Wu, Zhen Liu, and Yuan Yuan. 2025. Ethical dilemmas and the reconstruction of subjectivity in digital mourning in the age of AI: an empirical study on the acceptance intentions of bereaved family members of cancer patients.Frontiers in Digital Health7 (2025). doi:10.3389/fdgth.2025.1618169

    work page doi:10.3389/fdgth.2025.1618169 2025

  22. [22]

    Shyam Sundar

    Eun Go and S. Shyam Sundar. 2019. Humanizing chatbots: The effects of visual, identity and conversational cues on humanness perceptions. Computers in Human Behavior97 (2019), 304–316. doi:10.1016/j.chb.2019.01.020

    work page doi:10.1016/j.chb.2019.01.020 2019

  23. [23]

    Greenwald

    Anthony G. Greenwald. 1976. Within-subjects designs: To use or not to use?Psychological Bulletin83 (1976), 314–320. doi:10.1037/0033-2909.83.2.314

    work page doi:10.1037/0033-2909.83.2.314 1976

  24. [24]

    Rebecca Gulotta, William Odom, Haakon Faste, and Jodi Forlizzi. 2014. Legacy in the age of the internet: reflections on how interactive systems shape how we are remembered. InProceedings of the 2014 conference on Designing interactive systems (DIS ’14). Association for Computing Machinery, New York, NY, USA, 975–984. doi:10.1145/2598510.2598579

    work page doi:10.1145/2598510.2598579 2014

  25. [25]

    Tomasz Hollanek and Katarzyna Nowaczyk-Basińska. 2024. Griefbots, Deadbots, Postmortem Avatars: on Responsible Applications of Generative AI in the Digital Afterlife Industry.Philosophy & Technology37 (2024). doi:10.1007/s13347-024-00744-w

    work page doi:10.1007/s13347-024-00744-w 2024

  26. [26]

    Jack Holt, James Nicholson, and Jan David Smeddinck. 2021. From Personal Data to Digital Legacy: Exploring Conflicts in the Sharing, Security and Privacy of Post-mortem Data. InProceedings of the Web Conference 2021 (WWW ’21). Association for Computing Machinery, New York, NY, USA, 2745–2756. doi:10.1145/3442381.3450030

    work page doi:10.1145/3442381.3450030 2021

  27. [27]

    Jack Holt, Jan David Smeddinck, James Nicholson, Vasilis Vlachokyriakos, and Abigail C Durrant. 2024. Post-mortem information management: exploring contextual factors in appropriate personal data access after death.Human–Computer Interaction(2024), 1–36

    work page 2024

  28. [28]

    Geoffrey Keppel and Thomas D. Wickens. 2004.Design and Analysis: A Researcher’s Handbook(4th ed.). Pearson Prentice Hall, Upper Saddle River, NJ

    work page 2004

  29. [29]

    Keunwoo Kim, Minjung Park, and Youn-kyung Lim. 2021. Guiding Preferred Driving Style Using Voice in Autonomous Vehicles: An On-Road Wizard-of-Oz Study. InProceedings of the 2021 ACM Designing Interactive Systems Conference(Virtual Event, USA)(DIS ’21). Association for Manuscript submitted to ACM Designing Conversations with the Dead 19 Computing Machinery...

    work page doi:10.1145/3461778.3462056 2021

  30. [30]

    Dennis Klass, Phyllis Rolfe Silverman, and Steven L. Nickman. 1996. Continuing Bonds : New Understandings of Grief. https://api.semanticscholar. org/CorpusID:145474119

    work page 1996

  31. [31]

    Joel Krueger and Lucy Osler. 2022. Communing with the dead online: chatbots, grief, and continuing bonds.Journal of Consciousness Studies29, 9-10 (2022), 222–252

    work page 2022

  32. [32]

    Jonathan Lazar, Jinjuan Heidi Feng, and Harry Hochheiser. 2017. Interviews and Focus Groups. InResearch Methods in Human-Computer Interaction (2nd ed.). Morgan Kaufmann, Chapter 8, 187–228

    work page 2017

  33. [33]

    AI Afterlife

    Ying Lei, Shuai Ma, Yuling Sun, and Xiaojuan Ma. 2025. " AI Afterlife" as Digital Legacy: Perceptions, Expectations, and Concerns. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–18

    work page 2025

  34. [34]

    Nora Freya Lindemann. 2022. The Ethics of ‘Deathbots’.Science and Engineering Ethics28 (2022). doi:10.1007/s11948-022-00417-x

    work page doi:10.1007/s11948-022-00417-x 2022

  35. [35]

    Alexander Manevich and Shani Aluma. 2025. Artificial Continuing Bonds: A Proposed Framework for Responsible Integration of Generative AI in Clinical Thanatology.OMEGA-Journal of Death and Dying(2025), 00302228251372418

    work page 2025

  36. [36]

    David Maulsby, Saul Greenberg, and Richard Mander. 1993. Prototyping an intelligent agent through Wizard of Oz. InProceedings of the INTERACT ’93 and CHI ’93 Conference on Human Factors in Computing Systems(Amsterdam, The Netherlands)(CHI ’93). Association for Computing Machinery, New York, NY, USA, 277–284. doi:10.1145/169059.169215

    work page doi:10.1145/169059.169215 1993

  37. [37]

    Johanna Meurer, Christina Pakusch, Gunnar Stevens, Dave Randall, and Volker Wulf. 2020. A Wizard of Oz Study on Passengers’ Experiences of a Robo-Taxi Service in Real-Life Settings. InProceedings of the 2020 ACM Designing Interactive Systems Conference(Eindhoven, Netherlands)(DIS ’20). Association for Computing Machinery, New York, NY, USA, 1365–1377. doi...

    work page doi:10.1145/3357236.3395465 2020

  38. [38]

    Wendy Moncur and David Kirk. 2014. An emergent framework for digital memorials. InProceedings of the 2014 conference on Designing interactive systems (DIS ’14). Association for Computing Machinery, New York, NY, USA, 965–974. doi:10.1145/2598510.2598516

    work page doi:10.1145/2598510.2598516 2014

  39. [39]

    Brubaker

    Meredith Ringel Morris and Jed R. Brubaker. 2025. Generative Ghosts: Anticipating Benefits and Risks of AI Afterlives. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 536, 14 pages. doi:10.1177/00302228251372418

    work page doi:10.1177/00302228251372418 2025

  40. [40]

    2011.Fundamentals of qualitative research

    Johnny Saldana. 2011.Fundamentals of qualitative research. Oxford university press

    work page 2011

  41. [41]

    Conversation

    Anna Xygkou, Panote Siriaraya, Alexandra Covaci, Holly Gwen Prigerson, Robert Neimeyer, Chee Siang Ang, and Wan-Jou She. 2023. The "Conversation" about Loss: Understanding How Chatbot Technology was Used in Supporting People in Grief.. InProceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM, 1–15. doi:10.1145/3544548.3581154 Ma...

    work page doi:10.1145/3544548.3581154 2023