Afterglow: A Place-Based Memorial Ecology for AI-Mediated Pet Bereavement

1 Introduction

• Pet bereavement can be dismissed, lacks established mourning rituals, and may leave caregivers feeling pressure to justify the relationship’s significance.The relationship may endure even as daily acts of care end.Source: §1 · p. 1
• A symbolic Pet reply or platform reminder can claim authority to comfort, judge, intervene, or request attention, making remembrance feel like a condition of care.The paper calls this pressure coercive retention.Source: §1 · p. 1
• Afterglow is a mobile memorial world that separates private remembrance, human exchange, witnessing, and sparse symbolic Pet messages across independently accessible places.A formative survey with 57 respondents informed the first prototype, followed by six guided roundtable walkthroughs with 20 unique participants across two builds.Source: §1 · p. 2
• The inquiry asks what conflicting needs arise around connection, likeness, and protection, and how authority is distributed across places and roles.The study examines where that separation becomes unstable through iterative design and participant interpretations.Source: §1 · p. 2
• Legible Restraint requires limits on relational authority to hold across speaker identity, generation context, trigger logic, data use, and exit paths.Designing for Goodbye carries this requirement through dormancy, return, and departure, so continued use is not a condition of care.Source: §1 · p. 2

2.1 Pet Loss, Disenfranchisement, and Continuing Bonds

• Pet loss is disenfranchised grief, so memorials can recognize the relationship and loss while supporting continuing bonds and user-directed remembrance or disengagement.Continuing bonds can coexist with restoration and deliberate disengagement rather than requiring either attachment or letting go.Source: §2.1 · pp. 2–3

2.2 Social and Culturally Situated Memorial Interaction

• Memorial visibility and audience response can validate pet grief, while silence or unwanted exposure can repeat the social minimization that makes disclosure difficult.Audience conditions are culturally situated and shape witnessing, public grief, ritual, and authority.Source: §2.2 · p. 3

2.3 Interaction Modalities and Temporal Commitments in Pet Memorial Systems

• Afterglow combines private curation, human witnessing, lightweight world activity, and generated Pet language, asking whether it remains available through absence without demanding maintenance.The reviewed pet-specific systems typically foreground one primary form of encounter, such as embodied ritual, re-encounter, social response, or symbolic reflection.Source: §2.3 · pp. 3–4

2.4 Generated Speaker Identity and Relational Authority

• A generated Pet persona is built from caregiver-provided traces and may feel intimate, but it has no authority to remember, forgive, or judge on the animal’s behalf.The interface must identify its speaker role as inferred and constrain its claims.Source: §2.4 · p. 4
• Afterglow extends analysis from a generated Pet persona to an ecology of places, human peers, data transitions, and platform rules that distribute intimacy and authority.This shifts attention from the represented persona alone to how elements of the memorial ensemble shape encounters.Source: §2.4 · pp. 4–5

2.5 Relational Privacy and Shared Pet Memory

• Memorial data is relational: a caregiver’s photograph may depict other people, and a story may contain a family member’s private information.This makes the materials used to generate a persona relevant to relational authority and privacy.Source: §2.5 · p. 5
• Pet memorial authority extends beyond the uploader because caregivers may contribute or contest profile materials, while prior work describes digital-remains control as distributed.Source: §2.5 · p. 5
• Privacy spans the memorial lifecycle—from source and audience through transformation, downstream use, and retention—and includes contesting uses or leaving the system.Research on conversational data notes context-dependent sharing expectations [51], while AI-afterlife work links preservation, access, transfer, and dispossession [25].Source: §2.5 · p. 5

2.6 From Visible Boundaries to Legible Restraint

• Boundaries must be recognizable and actionable: silent blocks or generic errors can feel like rejection, while seamful design makes limitations visible.Source: §2.6 · p. 5
• Disclaimers may not affect trust, authoritative style can increase trust and persuasiveness, and chatbot endings can feel like loss shaped by agency and finality [40], motivating study of boundaries across responses and participation.Source: §2.6 · p. 5

3.1 Research Orientation and Evidence Boundaries

• The inquiry treats its findings as situated, intermediate-level design knowledge grounded in an artifact, and excludes therapeutic efficacy, grief reduction, clinical safety, and long-term behavior.The survey and walkthrough materials were primarily in Chinese; bilingual authors reviewed English translations against the original wording.Source: §3.1 · pp. 5–6, §3.2 · p. 6

3.3 Survey Instrument and Recruitment (𝑁= 57)

• The formative survey included 57 respondents, while each open-ended item received 51 responses; 32 respondents reported a companion-animal death.The instrument elicited reactions to virtual reunion, psychological support, AI-generated pet content, and AI moderation.Source: §3.3 · pp. 6–7

3.4 Formative Analysis and Initial Requirements

• The survey’s small cross-sectional groups suggested flexible interaction intensity, not a progression through grief stages.Deeper conversation or companionship was selected by 71% within one year of loss (n=7), 69% at one to three years (n=13), and 17% beyond three years (n=12); 75% in the last group selected observer companionship.Source: §3.4 · p. 7
• Respondents wanted specificity and recognizable likeness, while others preferred quiet observation and warned that human-like speech could feel false.Some described desired visual and vocal likeness, personality, and dialogue; anticipated benefits and risks could coexist in the same response record.Source: §3.4 · pp. 7–8
• Quiet observation and ordinary routines framed the memorial as a place for companionship as well as grief-focused support.Responses imagined a pet in a parallel world, quiet visits, or everyday care such as feeding, walking, and settling to sleep.Source: §3.4 · p. 8
• Survey responses paired the wish to talk with a pet with concern about emotional dependence, and imagined reunion alongside worries about prolonging pain.Other responses described potential comfort through memory preservation and acknowledgement, while also anticipating disappointment when generated likeness failed.Source: §3.4 · p. 8
• The initial requirements combined personally specific remembrance with caregiver-derived framing, voluntary return, and understandable AI generation, sharing, and protective intervention.Walkthrough conflicts within those requirements motivated the later tension families.Source: §3.4 · p. 8

4.1 Distributing the Memorial Relationship Across Places

• Afterglow distributes the bond across private remembrance, human witnessing, Pet-addressed messages, and shared memorial traces, making the generated Pet one bounded encounter.The Cottage supports mourner-directed remembrance, while the River, Post Office, and Forest offer different forms of exchange and witnessing.Source: §4.1 · p. 8

4.2 Role Distinctions Across Two Evaluated Builds

• Across both builds, generated roles could accumulate claims to continuing presence, lived experience, reassurance, and authority, while Pet-like prompts implied nearness despite target boundaries.First-person framing and sampled examples could confer authority that later boundary statements attempted to withdraw, making restraint an architectural concern.Source: §4.2 · pp. 9–10
• A proposed human-first MoMo path reserves the initial response opportunity for people, offering one short, explicitly requested AI acknowledgement only if no human reply arrives.A human reply keeps MoMo off by default, and the backstop identifies its artificial source.Source: §4.2 · p. 11
• Guardian moderates interpersonal conduct rather than certifying a Pet message’s truth or emotional safety, while the System role keeps functional authority distinct from Pet affection.The retained Guardian schema lacks a target field, so self-blame and interpersonal attack enter the same decision path.Source: §4.2 · p. 11

4.3 Privacy and Sharing Boundaries

• Consent to contribute, generate, and share are separable decisions, but one retained composer defaults to public rather than private-first.The research configuration stored app state locally, used a locally hosted model, and retained structured safety metadata instead of raw text, reducing third-party exposure.Source: §4.3 · p. 11
• Production still requires provenance, co-caregiver objection and removal paths, retention disclosure, verified deletion, and tested access control.The design treats consent to contribute, generate, and share as distinct, and calls for contestable permissions for shared memories.Source: §4.3 · p. 11

4.4 Temporal Openness and Designing for Goodbye

• Because visits, watering, and collection make activity measurable, repeated use may signify care, recognition, progress, or obligation; pause, export, close-this-memory, and return-later paths emerged as design directions.Afterglow keeps locations independently accessible and access persistent through absence; user-defined timelines, peer acknowledgement, social support, and intact dormancy are proposed criteria.Source: §4.4 · p. 12

4.5 Implementation and Scope

• The retained SwiftUI/Node.js prototype served gemma3:12b-it-q4_K_M locally; Postal and MoMo sampled four local few-shot examples with distinct temperature and token-cap settings.Postal used temperature 0.8 and a 220-token cap; MoMo used temperature 0.6 and a 100-token cap.Source: §4.5 · pp. 12–13
• Probabilistic prompt constraints, binary moderation ambiguity, and users’ agent attributions made the distribution of authority a central RtD concern.Source: §4.5 · p. 13
• The first-person Postal prompt positioned continuing pet presence more strongly than the analytic target boundary, bringing role separation into the inquiry.The study traced how the provisional artifact distributed and contested authority.Source: §4.5 · p. 13

5.1 Participants and Roundtable Procedure (𝑁= 20)

• Six 60-minute roundtables included 20 unique participants and 30 attendances across an initial build and a revised demo.Ten participants used the initial build; all returned for Round 2, which also included ten newly recruited participants.Source: §5.1 · p. 13
• The facilitator used a shared end-to-end walkthrough and three probes about Pet absolution, AI acknowledgement, and intervention against directed hostility.Roundtable discussion, rather than controlled build-version effects or within-person change, was the analytic focus.Source: §5.1 · p. 14

5.2 Designer-Led Interpretive Analysis and Coding Traceability

• The interviewer-designer coded contemporaneous notes at the level of comments and discussion episodes, tracing open codes into focused codes and design interpretations.The codebook and artifact history support traceability; one additional code remained provisional and outside the three reported tension families.Source: §5.2 · pp. 14–15, Figure 4 · p. 15
• The analysis retained counterpositions: desires for reassurance and acknowledgement coexisted with doubts about machine judgment, solitude, and human witnessing.The three tension families synthesize conflicts rather than rank attitudes toward AI or estimate how many sessions endorsed each position.Source: §5.2 · p. 15

5.3 Traceability from Evidence to Design

• Table 4 links the three tension families to design responses and records their rationale.The table distinguishes the initial build (E1), revised demo (E2), and proposed production mechanisms (P).Source: §5.3 · p. 15

6 Findings

• Three design tensions emerged as salient in walkthroughs and roundtables: returnable connection versus emotional obligation, recognizable likeness versus ontological clarity, and protective intervention versus surveillant authority.Table 4 traces the three tension families to design responses.Source: §6 · p. 15, Table 4 · p. 16

6.1 Theme 1. Returnable Connection versus Emotional Obligation

• Participants valued social recognition of pet grief and small acknowledgements, while accounts also positioned human witnessing as complementary to a brief MoMo reply.Seeing another person mourn could matter more than a bot’s reassurance, while a small reply could help when no human response was present.Source: §6.1 · p. 16
• Private remembrance and public expression served different needs: the Cottage protected solitude, while shared spaces enabled witnessing but made audience and response consequential.One account described writing as a way to live with loss rather than seek sympathy.Source: §6.1 · p. 16
• Watering, tree growth, reactions, and collection offered lightweight reasons to return, but visible accumulation could make remembrance look like participation that must be performed.Participants also imagined brief, voluntary visits for private company and everyday sharing with the represented pet.Source: §6.1 · pp. 16–17

6.2 Theme 2. Recognizable Likeness versus Ontological Clarity

• Desired likeness depended on specific appearance, voice, habits, personality, and routines, while adult-like life advice could make the Pet seem like a chatbot using the pet as a mask.Participants described such advice as breaking the illusion of the represented cat.Source: §6.2 · p. 17
• Participants preferred visiting a represented pet with its own activity and invitations, while wished-for scenes were more comfortable in a bounded memorial place than in everyday surroundings.Examples included filling a missing childhood-pet picture and imagining a lost dog meeting a living dog in a memorial space.Source: §6.2 · pp. 17–18
• Likeness can support recognition, but life advice or claims of continuing presence alter what the Pet appears entitled to say; vivid representation therefore needs a bounded speaker role.Participants’ preference for realistic appearance and concerns about artificial authority addressed different aspects of representation.Source: §6.2 · p. 18

6.3 Theme 3. Protective Intervention versus Surveillant Authority

• Pet functions such as organizing material or producing short sensory messages fit a bounded expressive role, whereas life advice and moral judgment shifted authority toward a therapist-like agent.In the staged euthanasia scenario, desire for reassurance coexisted with concern about unqualified Pet absolution.Source: §6.3 · p. 18
• Participants supported intervention against harm to others but resisted moderation that monitored grief or blocked anger, guilt, and self-blame.Protection was distinguished by the target of an expression, preserving emotional intensity as part of grief expression.Source: §6.3 · p. 18

7.1 From Responsive Memorials to Returnable Memorial Relations

• Transitions change who can respond, what memory material can become, and what participation appears to require, making interactional demand—not only medium—a design concern.A Cottage memory can become apparent Pet speech in a Postal reply, while sharing in the River changes its audience.Source: §7.1 · p. 19
• Memorial relations depend on how much participation a return asks of mourners: visits, quiet expression, human witnessing, and Pet welfare independent of attention allow connection on different terms.Participants distinguished visiting from responsibility for the represented Pet’s welfare; a short visit can carry meaning while silence remains private.Source: §7.1 · p. 19
• Participants distinguished private expression, human presence, and visiting from seeking sympathy, bot reassurance, and responsibility for the Pet; combining these roles in one chatbot could blur their authority.The discussion motivates an opt-in MoMo acknowledgement after a human-first response window.Source: §7.1 · p. 19

7.2 Legible Restraint as Cross-Layer Relational Authority

• Legible Restraint requires announced boundaries to persist through transformations, responses, and returns by coordinating generation context, trigger logic, data use, and exit paths.The requirement concerns contradictions across layers, not disclosure alone.Source: §7.2 · p. 20
• MoMo’s disclosed AI label did not prevent prompts and examples from invoking continuing presence, moral reassurance, and lived loss, so interface copy alone was insufficient.The analysis calls for changing generation context as well as interface copy.Source: §7.2 · p. 20
• The retained moderation gate treated self-blame and directed hostility alike, motivating target-sensitive review, an explanation and revision path, and a response window before requested AI acknowledgement.The proposed design separates human response from explicitly requested AI response.Source: §7.2 · p. 20

7.3 Designing for Goodbye as a Temporal Consequence

• Countable resources can make witnessing tangible but may turn remembrance into maintenance debt, so visits should leave optional traces while resources remain separate from welfare, persistence, and access.The proposed longitudinal evaluation would pair peer outcomes with activity traces and examine whether activity feels like welcome witnessing or maintenance debt.Source: §7.3 · pp. 20–21
• Dormancy preserves a memorial, re-entry restores it intact, export separates memory keeping from service use, and user-initiated departure leaves emotional closure to the mourner.Transitions should preserve the authority boundaries of an active visit.Source: §7.3 · p. 21

7.4 Following the Boundary Through Shared Data

• Shared memories can create conflicts when caregivers prefer different transformations or audiences, but the prototype assigns contribution, generation, and sharing decisions to one uploader.The discussion presents caregiver disputes as a future empirical question and proposes linked histories with separate permissions for preservation, generation, and sharing.Source: §7.4 · p. 21

7.5 Why Pet Bereavement Sharpens the HCI Problem

• Pet bereavement combines limited social recognition with an authorization gap for generated speech, so greater social and generative responsiveness can validate mourners while overstating a Pet’s authority.Afterglow positions human peers, symbolic Pet speech, and platform governance differently within a place-based ecology.Source: §7.5 · p. 21

7.6 Cultural Situatedness of Memorial Architecture

• The findings are culturally situated: roundtable participants identified as Asian, were 18–35, and sessions were primarily in Chinese; cross-context research remains open.The discussion calls for examining role boundaries across mourning practices and inviting participant-authored alternatives to the world’s metaphors.Source: §7.6 · p. 22

Limitations and Future Work

• The samples were demographically concentrated: women were 46 of 57 survey respondents and 17 of 20 roundtable participants; all roundtable participants identified as Asian and were 18–35.The sessions were conducted primarily in Chinese, situating the findings culturally.Source: §8 · p. 22
• Written notes were the sole session record, and the first author served as designer, facilitator, and coder, limiting turn-by-turn reconstruction, prevalence claims, and independent interpretation.The notes support discussion-level tensions and links to prompts, interface transitions, and artifact versions.Source: §8 · p. 22
• Repeated participation added continuity to formative evaluation but was distinct from living with the system between sessions; build revision, prior exposure, and group discussion leave individual changes unresolved.The six hour-long walkthroughs included ten returning participants and ten newly recruited participants encountering the revised build.Source: §8 · p. 22
• The authors propose a six- to eight-week multi-user deployment combining activity logs, experience-sampling prompts, and follow-up interviews to examine everyday use, absence, and exit.They also suggest participant transcript review and opt-in recording when ethically appropriate.Source: §8 · p. 22

9 Conclusion

• Limits must align symbolic expression, human judgment, and voluntary return across the encounter, including dormancy, return, and departure.The aim is remembrance without making continued use a measure of care.Source: §9 · pp. 22–23
• Participants wanted recognizable remembrance, quiet visits, voluntary human witnessing, and room to step away; retained-role mismatches showed visible separation remained incomplete.The design tests traced these tensions across generation, shared interaction, and absence.Source: §9 · pp. 22–23
• The evidence is situated in a primarily Chinese-language, designer-led formative study; sustained use and proposed production mechanisms need further evaluation.Longitudinal and cross-cultural work can examine whether boundaries remain meaningful as memorial practices change.Source: §9 · p. 23

A English Translation of the Formative Survey

• The translated appendix presents the 24-item survey and reports a qualitative corpus of three open-ended items, each with 51 responses.The survey was administered in Chinese, and “Other” denotes a free-text option.Source: A English Translation of the Formative… · p. 26

A.1 Part 1: About You

• The participant section asks demographic and pet-history questions; overlapping age-bin boundaries at 30 and 40 mean no age-group comparisons are reported.The original response bins overlapped at those values.Source: §A.1 · p. 26

A.2 Part 2: Pets and Emotions; A.3 Part 3: Virtual Experience Ideas

• The survey asks about social recognition and support for pet loss, memory practices, and preferences or concerns about virtual reunion, sharing, AI messages, and moderation.It also asks about preferred formats and desired features.Source: §A.2 · p. 26, §A.3 · pp. 26–27

B Retained Model and Role Configuration

• Appendix B documents retained backend settings and the prototype’s generative range, including locally sampled examples for each Postal or MoMo request.The appendix also distinguishes deterministic system messages from the three LLM roles.Source: B Retained Model and Role Configuration · p. 27, Table 5 · p. 28, §B.1 · p. 28

B.1 System Instructions

• The role instructions specify a comforting Pet reply, a Guardian moderation judgment, and MoMo’s identity as a nonhuman digital guardian.The Pet role is prompted to reply in the pet’s voice; MoMo is instructed to acknowledge having no body.Source: §B.1 · p. 27

• Postal examples reassured caregivers they had done nothing wrong and depicted the pet as present, conflicting with the later prohibition on moral absolution.Source: §B.2 · p. 28
• MoMo examples framed MoMo as a bereaved real friend and concealed its AI identity, conflicting with the target of explicit AI identity.Source: §B.2 · p. 28
• Because sampled examples entered generation context, they could override intended distinctions even when the main prompt stated a different rule.Source: §B.2 · p. 28

• The analytic map supports conceptual traceability from contemporaneous interviewer notes and discussion episodes, but not an audio-verified transcript or participant-frequency matrix.Source: C Roundtable Coding Map and Full… · p. 28
• A small reply could make someone feel noticed when nothing happened, while seeing another person miss their dog was described as helping more than a bot.Source: C Roundtable Coding Map and Full… · p. 29
• One participant imagined brief visits to check for a new message and water the tree rather than staying long.Source: C Roundtable Coding Map and Full… · p. 29
• A cat giving life advice could feel like a chatbot pretending to be the pet, breaking the illusion.Source: C Roundtable Coding Map and Full… · p. 29
• One participant wanted to see that their pet was okay without continuing to assign it tasks.Source: C Roundtable Coding Map and Full… · p. 29
• A participant saw generated imagery as filling in a wished-for picture with a childhood pet they had never taken together.Source: C Roundtable Coding Map and Full… · p. 29
• Mixed reality with a lost dog in real surroundings could feel creepy, though one participant still wanted a place for the lost and living dogs to meet and play.Source: C Roundtable Coding Map and Full… · p. 29
• Another participant accepted an imperfect Pet voice if it avoided acting like a therapist and remained recognizably their pet.Source: C Roundtable Coding Map and Full… · p. 29