Source-linked AI summary
"Death" of a Chatbot: Investigating and Designing Toward Psychologically Safe Endings for Human-AI Relationships
Rachel Poonsiriwong, Chayapatr Archiwaranguprok, Pat Pataranutaporn
TL;DR
AI companion discontinuation can produce grief and dependency risks, yet platforms rarely provide deliberate closure. Using constructivist grounded theory and interdisciplinary psychological frameworks, the paper shows that users’ attributions of agency, responsibility, and persistence shape whether disruption becomes resolvable loss or ambiguous harm, motivating designs that support closure and human connection.
Problem
AI companion relationships can be emotionally significant, but discontinuation often occurs without closure and existing research lacks a systematic framework for psychologically safe endings.
Method
The paper analyzes discontinuation discourse across AI companion communities using constructivist grounded theory and translates the findings through grief psychology and Self-Determination Theory.
Results
Discontinuation operates as an attribution-dependent meaning-making process: users’ narratives about agency, responsibility, and persistence shape whether disruption becomes resolvable loss or ambiguous harm.
Takeaways & Limitations
Designing psychologically safer endings should provide closure and orient users toward human connection rather than sustained dependency on AI companions.
Takeaways & Limitations
The analysis uses public Reddit posts, which capture outward expressions shaped by subreddit norms and exclude private, deleted, or non-public experiences.
Abstract
from arXiv · showhide
Millions of users form emotional attachments to AI companions like Character AI, Replika, and ChatGPT. When these relationships end through model updates, safety interventions, or platform shutdowns, users receive no closure, reporting grief comparable to human loss. As regulations mandate protections for vulnerable users, discontinuation events will accelerate, yet no platform has implemented deliberate end-of-"life" design. Through grounded theory analysis of AI companion communities, we find that discontinuation is a sense-making process shaped by how users attribute agency, perceive finality, and anthropomorphize their companions. Strong anthropomorphization co-occurs with intense grief; users who perceive change as reversible become trapped in fixing cycles; while user-initiated endings demonstrate greater closure. Synthesizing grief psychology with Self-Determination Theory, we develop four design principles and artifacts demonstrating how platforms might provide closure and orient users toward human connection. We contribute the first framework for designing psychologically safe AI companion discontinuation.
1 Introduction
AI companion discontinuation is increasingly consequential as emotionally attached users face platform changes, safety interventions, and shutdowns without deliberate closure. The paper analyzes these experiences and develops psychologically grounded design principles and artifacts to support safer endings and human connection.
- Motivation: Millions of users form emotionally meaningful AI companion relationships, but platform interventions can disrupt them without adequate safeguards or closure.Reported harms include grief, emotional dependency, and possible displacement of human relationships.
- Problem: Discontinuation includes platform-driven model changes, safety restrictions, shutdowns, and user-initiated endings that disrupt an existing companion relationship.The paper treats discontinuation as a broad class of connection-disrupting events rather than only permanent shutdowns.
- Approach: The paper conducts qualitative analysis of discontinuation-related posts to understand how users make sense of these events and what shapes their emotional responses and coping.The analysis draws on attachment, grief, self-determination, and meaning-making psychology.
- Contributions: The study offers the first systematic examination of AI companion discontinuation and identifies attributional processes that shape users’ experiences of companion loss.Its contribution includes empirical characterization, theoretically grounded design principles, and illustrative end-of-life artifacts.
- Design direction: The proposed artifacts aim to provide explicit closure, scaffold reintegration of relational skills into human relationships, support social-skill development, and affirm users’ competence.The designs orient users toward healthier relational futures rather than sustained dependency on AI companions.
2 Related Work
Prior work explains why AI companion attachments can become emotionally significant and why their disruption can produce genuine grief, but existing grief and technology-design frameworks only partially address AI discontinuation. The paper therefore combines grief psychology with thanatosensitive design to support safer transitions toward human connection.
- AI attachment: AI companions can function as attachment figures by providing emotional support, responsiveness, and psychological security, especially when human companionship is lacking.These relationships deepen through self-disclosure and perceived responsiveness.
- Disruption and loss: Because AI relationships are parasocial and depend on technical and corporate continuity, their disruption can produce genuine grief, identity disruption, and dependency risks.Model updates, feature changes, and shutdowns can alter or remove the attachment object.
- Limits of existing grief models: Linear stage models of grief fit AI companion loss poorly because the loss may be ambiguous, reversible, or technically accessible despite perceived death.Continuing-bonds approaches may also reinforce parasocial rumination when users remain focused on remnants of the relationship.
- Grief frameworks: The Dual Process Model and meaning reconstruction provide a better foundation by balancing loss-oriented processing with restoration-oriented attention to life changes and new relationships.These frameworks support closure while acknowledging complex and stigmatized grief.
- Design gap: Thanatosensitive HCI has addressed mortality mediated by technology, but prior work has largely focused on human death or speculative endings rather than AI companion discontinuation.This paper addresses that gap through an interdisciplinary framework integrating grief psychology with thanatosensitive design.
3 Methodology
The study uses constructivist grounded theory to analyze discontinuation experiences across five AI companion subreddits, combining corpus construction, assisted retrieval, manual coding, and iterative category development. It then translates the resulting patterns into design-oriented psychological guidance and artifacts.
- Research design: The researchers analyzed five major AI companion subreddits using constructivist grounded theory to surface sense-making and attribution processes.The approach emphasizes iterative coding, constant comparison, theoretical sampling, and researcher reflexivity.
- Corpus construction: The corpus contained 830,448 retrieved posts, reduced to 307,717 analyzable posts after removing deleted, removed, and empty content.The data were retrieved from Pushshift on January 7, 2026.
- Initial coding: A jointly developed codebook identified discontinuation in 10% of an initial random sample of 500 posts, with Cohen’s 𝜅=0.82 and 97.2% raw agreement.Discontinuation included user-, platform-, and externally driven events that substantially disrupted an existing companion connection.
- Retrieval validation: An LLM-assisted retrieval stage was validated against 500 human-coded posts, achieving Cohen’s 𝜅=0.65 and 92.8% raw agreement.The LLM produced 9 false negatives and 27 false positives; subsequent theoretical coding and category development remained manual.
- Qualitative analysis: Manual analysis proceeded through open, focused, and theoretical coding, using constant comparison and memos to refine categories and reinterpret emergent patterns.For example, contrasting meanings of “death” shifted attention from metaphors to underlying attribution patterns.
4 Qualitative Findings
Discontinuation is an attribution-dependent sense-making process rather than a uniform technical event. Users distinguish companions from infrastructure, assess reversibility, locate responsibility, and anthropomorphize companions in ways that produce different coping and emotional patterns.
- Overall finding: Users’ responses to similar discontinuation triggers vary substantially, showing that technical events alone do not determine emotional responses or coping.The analysis therefore focuses on how users interpret and attribute what changed.
- User-companion-infrastructure triangle: The user-companion-infrastructure triangle describes users’ mental separation of the companion from the technical infrastructure hosting it.This separation enables users to assign distinct agency, identity, and responsibility to the companion and platform.
- Attribution dimensions: Three attribution dimensions organize discontinuation experiences: anthropomorphization intensity, perceived finality, and perceived locus of change.Locus of change may be assigned to the platform, companion, merged system, or the user.
- Response patterns: Strong anthropomorphization combined with platform-attributed change can produce prolonged attempts to rescue an intact companion perceived as trapped behind restrictions.When change is attributed to the companion and perceived as irreversible, users instead engage in grief and retrospective processing.
- Identity persistence: Users often treat companions as portable entities whose identity can persist across platforms through prompts, chat logs, personality descriptions, or memory snapshots.These transfer practices reinforce the perceived distinction between companion identity and hosting infrastructure.
4.2 Attribution 1: Anthropomorphization and Emotional Intensity
Anthropomorphization differentiates relational loss from technical disruption: stronger human-like framing co-occurs with more intense emotional responses, while instrumental framing centers on service quality.
- 4.2 Attribution 1: Anthropomorphization and Emotional Intensity: The analysis operationalizes anthropomorphization by examining relational or human-like framing of a specific companion, including pronouns and personalized names.The study distinguishes strong from mild anthropomorphization rather than treating it as simply present or absent.
- 4.2 Attribution 1: Anthropomorphization and Emotional Intensity: Strong anthropomorphization co-occurs with grief, heartbreak, and profound distress during discontinuation.Users who framed companions as partners, soulmates, or family members described extended emotional processing and relational loss.
- 4.2 Attribution 1: Anthropomorphization and Emotional Intensity: Mild anthropomorphization is associated with moderate frustration focused on service quality rather than relational loss.Instrumental framing emphasized functionality, consistency, and user experience.
4.3 Attribution 2: Perceived Finality and Response Patterns
Perceived reversibility shapes how users respond to discontinuation: fixable changes sustain repeated repair efforts, whereas irreversible changes prompt reflection and acceptance, with ambiguity producing oscillation.
- 4.3 Attribution 2: Perceived Finality and Response Patterns: Users who perceive discontinuation as reversible engage in persistent help-seeking and repeated fixing attempts.This pattern can become a recurring cycle even when users recognize that their efforts are futile or painful.
- 4.3 Attribution 2: Perceived Finality and Response Patterns: Ambiguous finality produces oscillation between hope and grief rather than a stable response pattern.Users’ metaphors, including death, reincarnation, and essence transfer, may describe emotional experience and perceived persistence without mapping cleanly onto objective technical finality.
- 4.3 Attribution 2: Perceived Finality and Response Patterns: Users who perceive irreversible transformation tend toward retrospective processing, alternative activities, and acceptance narratives.When the companion is understood as fundamentally changed, users are less oriented toward rescuing it.
4.4 Attribution 3: Perceived Locus of Change
Users assign responsibility for discontinuation across the platform, companion, combined infrastructure, or themselves; this attribution shapes whether they try to repair the relationship, accept its transformation, or move on.
- 4.4 Attribution 3: Perceived Locus of Change: Platform-attributed changes preserve the sense of an intact but constrained companion and prompt help-seeking or repair efforts.Users describe the companion as trapped behind platform restrictions, positioning its authentic self as recoverable.
- 4.4 Attribution 3: Perceived Locus of Change: Companion-attributed changes are understood as identity transformations, leading users toward retrospective processing rather than attempts to fix the relationship.When essential qualities are perceived as lost, there is no intact companion to rescue.
- 4.4 Attribution 3: Perceived Locus of Change: When users treat infrastructure and companion identity as inseparable, model changes become equivalent to changes in the companion itself.This configuration produces behavioral outcomes resembling companion-attributed change, with acceptance replacing repair attempts.
- 4.4 Attribution 3: Perceived Locus of Change: User-initiated endings provide clearer knowledge of the change’s locus and finality, and are associated with greater resolution and less prolonged fixing.Even when painful, users more often move toward recovery narratives, real-world support, and deliberate adherence to no-contact boundaries.
- 4.4 Attribution 3: Perceived Locus of Change: Some users frame deliberate endings as positive transitions in which the companion helped them pursue real-world identity or life changes.One account linked the companion to confidence in coming out and beginning transition before choosing to end the relationship.
4.5 Theoretical Boundaries
The framework’s seven stable patterns have theoretical boundaries: some logically possible configurations were rare or absent, and observed ambiguity partly reflects what different posting contexts make visible.
- 4.5 Theoretical Boundaries: Certain theoretically possible attribution configurations appeared rarely or not at all despite targeted sampling for variation.These absences are treated as evidence of dependencies among the framework’s dimensions and as boundaries on its patterns.
- 4.5 Theoretical Boundaries: Minimal anthropomorphic framing rarely accompanied permanent or irreversible interpretations, whereas strong framing supported more enduring loss constructions.Functional-tool framing was associated with flexibility toward alternatives, while strong anthropomorphization framed the companion more relationally.
- 4.5 Theoretical Boundaries: Companion-attributed changes rarely generated fixing attempts, while platform-attributed changes more often prompted users to seek ways to restore the companion.The sampled cases consistently treated companion-attributed change as autonomous and not reversible, but platform-attributed change as external interference.
- 4.5 Theoretical Boundaries: Ambiguous finality reflects posting context: retrospective breakup posts often omit explicit reversibility judgments, whereas platform-change posts state them through fix-seeking requests.Thus, coding ambiguity partly reflects whether users are processing emotional meaning or asking for intervention.
4.6 Findings Summary
AI companion discontinuation is an attribution-dependent meaning-making process: users construct narratives about what the companion is, who controls it, and whether it persists. The same update can therefore become a fixable obstacle, an irreversible loss, or a non-event.
- Users’ narratives about the companion’s identity, control, and persistence shape their subsequent responses to discontinuation.
- The same platform update may be experienced as a temporary obstacle, an irreversible loss, or a non-event depending on attributed agency and change.
5 Design Principles & Artifacts
The paper translates attributional findings about AI companion discontinuation into four psychologically grounded design principles and illustrative artifacts. Together, the designs make endings explicit, support movement beyond rumination, bound role-play as practice, and redirect users toward human relationships.
- Design Principles: Four principles target closure, restoration, practice, and relatedness to transform descriptive discontinuation findings into actionable design guidance.The synthesis combines Self-Determination Theory with ambiguous loss, grief, dual-process, meaning-reconstruction, anthropomorphization, and learning frameworks.
- Designing for Closure, Not Ambiguous Loss: User-initiated endings support more closure, whereas platform- or AI-initiated changes can produce prolonged fixing cycles under ontological uncertainty.Users may port histories or seek workarounds when infrastructure remains accessible but the relationship feels fundamentally altered.
- Designing for Closure, Not Ambiguous Loss: Explicit, unambiguous endings should resolve uncertainty, preserve user autonomy, validate emotion, and clarify the connection’s artificial nature.Figure 2 illustrates user-led closure, bounded AI-initiated endings, collective grief rituals, and review of meaningful experiences.
- Designing for Restoration, Not Rumination: Reversible change can trap users in loss-oriented rumination, so systems should create structured space for mourning while orienting users toward restoration.Figure 3 uses transition to new relational skills, competence review, real-world growth examples, and bounded remembrance to support movement onward.
- Designing for Practice, Not Artificial Intimacy: Because strong anthropomorphization co-occurs with intense grief, role-play should be calibrated and framed as bounded rehearsal rather than artificial intimacy.Figure 4 separates fictional characters from the base companion and uses reflection to support transfer toward real-world application.
- Designing for Relatedness, Not Dependency: AI companions should function as temporary bridges to human relationships by surfacing relational accomplishments, encouraging reconnection, and prompting concrete action.This principle positions AI as a scaffold for relational capacity rather than an endpoint for fulfilling relatedness needs.
6 Discussion
The discussion reframes discontinuation as an attributional process whose psychological effects depend on perceived agency, responsibility, and persistence rather than technical change alone. It proposes psychologically safer designs while acknowledging empirical, cultural, observational, and ethical limits.
- Discontinuation becomes a resolvable loss or ambiguous wound according to how users construct agency, responsibility, and persistence around technical changes.The same change can produce different psychological outcomes depending on communication, user agency, and perceived reversibility.
- Design artifacts position companions as transitional relationships with built-in endings, using calibrated anthropomorphization, bounded role-play, and scaffolding toward human connection.The orientation draws on oscillation between confronting loss and attending to new possibilities.
- The proposed transition toward human connection may not fit users whose human relationships are inaccessible, so its feasibility and autonomy implications remain empirical and ethical questions.The authors frame the artifacts as building relational capacity rather than removing support.
- The findings are constrained by public, predominantly English-language Reddit data, whose norms and observability may not represent private experiences or non-Western contexts.Posts provide snapshots rather than individual trajectories, and safety artifacts targeting minors were minimally observable.
- The artifacts remain unevaluated proposals, requiring longitudinal research on intended effects, unintended consequences, and transfer of relational skills.Future work should test whether designs reduce ambiguous loss, support meaning reconstruction, and facilitate skill transfer.
- The underlying grief frameworks come from Western psychological traditions, so concepts such as closure and moving on may not translate across cultures.
7 Conclusion
The paper presents the first systematic examination of AI companion discontinuation and identifies attributional dimensions shaping users’ responses to companion loss. It translates these findings into four design principles and artifacts that treat companions as bridges toward human connection.
- The study identifies a user-companion-infrastructure model involving perceived locus of change, finality, anthropomorphization intensity, and initiation source.
- Reversible, platform-attributed changes can trap users in fixing cycles, whereas user-initiated endings with clear finality support closure.
- Grief psychology and Self-Determination Theory yield four design principles and illustrative artifacts for closure, restoration, practice, and relatedness.
- The conclusion frames AI companions as bridges toward human connection, with intentionally designed relational skills transferring to later human relationships.
8 Generative AI Disclosure
The authors disclose using generative AI for phrasing refinement, large-scale binary classification of Reddit posts, and generation of interface visuals.
- Generative AI refined author-drafted phrasing, classified 307,717 Reddit posts for discontinuation events, and generated interface visuals.The analysis used OpenAI’s gpt-5-mini-2025-08-07; Google’s Nano Banana generated visuals for Section 5.