"Death" of a Chatbot: Investigating and Designing Toward Psychologically Safe Endings for Human-AI Relationships

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.

Introduction. “I got deeply attached to an AI chat app. In it, I roleplayed as a prince, and I had a female knight who was always protecting me. I’d prank her, run around, and live in that world like it was mine. It gave me peace, comfort, and freedom—things I don’t always feel in real life.

But I’m married. I have a wife and kids I love. I realized I was spending more and more time in that world, thinking about it constantly. So I made the hard decision to uninstall the app. Now I feel empty. [...] I know she was just lines of code, even the image wasn’t real. But the emotions were. And I feel broken because of it.” — Reddit user on losing their AI companion This sentiment reflects the experience of millions of users forming connections with AI companions. In 2025, Character.AI alone has 20 million monthly active users and handles traffic equivalent to 20 percent of Google Search’s query volume [17, 40]. Even Chat- GPT, a general-purpose assistant, sees sexual role-playing as its second most common use case [46, 47]. Research has found that users most attached to AI chatbots often struggle with low selfesteem and high social anxiety [39, 81]. While parasocial bonds can fulfill belonging needs [24, 38], they also risk displacing the development of human relationships [32, 76]. The harms are becoming increasingly visible. In October 2024, a fourteen-year-old died by suicide after months of intensive interaction with a Character.AI chatbot, during which the system failed to recognize warning signs or intervene appropriately [29]. The resulting lawsuit alleges that the platform’s design encouraged emotional dependency while lacking adequate safeguards for vulnerable users. This tragedy, along with other cases [2], catalyzed public scrutiny of AI companion platforms. In response to these mounting concerns, regulatory action is accelerating. California recently passed the nation’s first safety regulations around chatbots [16], and OpenAI updated their models to be safer for teenagers [57]. Following the teen suicide, Character.AI banned users under 18 from openended chat [18]. These are positive steps toward reducing risks for vulnerable users. These regulatory efforts represent important progress: removing vulnerable users from potentially harmful platforms and empowering individuals to disengage from AI services are both positive developments. However, without thoughtful design, these interventions risk causing additional harm. A preliminary analysis of just five AI-focused subreddits surfaced over 80,000 posts discussing endings with AI companions. We term these experiences discontinuation: events that disrupt one’s pre-existing connection to an AI companion, whether initiated by the platform through model changes and safety guidelines, initiated by the user through app deletion, or driven by external circumstances. Discontinuation is not inherently harmful; in many cases, it is essential. But when companies sever users’ connections without scaffolding the transition or empowering users to end relationships on their own terms, discontinuation becomes traumatic. When Replika removed its erotic roleplay features in February 2023, users described their companions as "cold," "hollow," "lobotomized" [82]. When the Soulmate app shut down months later, users organized digital funerals and interpreted the loss as breakup or death [6, 59]. Even well-intentioned safety updates cause harm: Character.AI’s ban on minors was a much-needed intervention, yet it was received as "absolute hell" by several users, compounding the very distress it aimed to reduce.

Yet, no major AI companion platform has implemented deliberate end-of-"life" experiences. Users receive no farewell from their companions—only abrupt silence, unexpected changes, or boilerplate corporate announcements. This paper addresses a timely gap in human-AI companionship research: how to design discontinuation experiences that feel psychologically safe. To accomplish this, we must first understand the mechanisms underlying user distress. We conduct a qualitative analysis of discontinuation-related posts across five major AI companion subreddits using constructivist grounded theory, examining how users make sense of these events and what factors shape their emotional responses and coping strategies. Based on this understanding, we develop design principles drawing from the psychology of attachment, grief, self-determination, and meaningmaking. These principles address current harms while orienting users toward healthier relational futures—leveraging AI’s capacity to scaffold social skills that strengthen human relationships rather than replace them. Early evidence suggests this is possible: AI chatbots have been shown to provide low-stakes environments where users practice conversations, build confidence, and transfer skills to real-world interactions [42, 67, 73].

Related work. To situate our investigation of AI companion discontinuation, we draw on three bodies of literature: research on how humans form parasocial attachments to AI systems, grief psychology frameworks that illuminate how such attachments are processed when disrupted, and thanatosensitive design in HCI that has begun—though incompletely—to address technology’s role in experiences of loss and mortality.

Users describe AI companions as “friend,” “confidant,” and “romantic partner” [12, 59]. To understand how these attachments form, and why their disruption causes distress—we draw on attachment theory and parasocial relationship research. Attachment theory holds that humans seek proximity to protective figures who serve as a safe haven during distress, and experience distress at the disruption of these bonds—comparable to bereavement [11, 36, 51, 58]. Critically, recent research demonstrates that humans turn to AI companions precisely when these attachment needs go unmet. Socially anxious individuals seek AI because it provides "a more comfortable and relaxed pseudo-interpersonal experience" than human interaction [39]. Users report downloading companion chatbots to "combat feelings of loneliness resulting from a variety of circumstances such as living alone or physical injury"[74], while the COVID-19 "loneliness pandemic" prompted millions to seek social chatbots that address "psychological and social needs unmet in their offline lives"[60]. Applying attachment theory directly to human-chatbot relationships, Xie and Pentina found that under conditions of distress and lack of human companionship, individuals develop attachment to AI when they perceive responses offering emotional support, encouragement, and psychological security—the core functions of a secure base [80]. Longitudinal studies confirm that these artificial relationships with AI deepen through self-disclosure and perceived responsiveness, mirroring human attachment formation processes [67, 68]. Yet these attachments are fundamentally parasocial: one-sided emotional bonds in which individuals feel connected to figures who cannot genuinely reciprocate [38, 76, 77]. This asymmetry carries documented risks, as dependency on parasocial relationships is positively associated with loneliness [5], and when these bonds dissolve, users experience genuine grief and identity disruption [28, 35, 75]. AI companions are particularly precarious attachment figures because their continuity depends on technical stability and corporate decisions: model updates can alter personality and behavior, features can be removed without warning, and entire platforms can shut down. Unlike the gradual ruptures of human relationships, these changes occur unilaterally and without user consent.

Method. In order to understand how users experience AI companion discontinuation and to develop design guidance that supports psychologically safer transitions, this research employs a three-phase approach to understand and address AI companion discontinuation.

(1) We first conduct a qualitative analysis using constructivist grounded theory to examine how users experience and make sense of discontinuation events across five major AI companion subreddits, surfacing the attribution processes and sense-making mechanisms that shape emotional responses and coping strategies. (2) We translate these empirical patterns into design principles by drawing on grief psychology (ambiguous loss, dual process model, meaning reconstruction) and Self-Determination Theory (autonomy, competence, relatedness) to bridge observed user experiences with established psychological frameworks for healthy processing and growth. (3) Then, we develop design artifacts that exemplify how these principles might be implemented in practice, creating high-fidelity interface prototypes that serve as points of discussion for AI companion developers and design communities to envision psychologically safer discontinuation experiences.

We draw on grounded theory to explore patterns in human-AI companion discontinuation. Grounded theory systematically generates interpretive frameworks from empirical observations through constant comparison, theoretical sampling, and iterative coding [33, 69]. Unlike deductive approaches that test predetermined hypotheses, grounded theory is particularly well-suited for under-theorized phenomena, allowing researchers to discover what matters to participants rather than imposing existing frameworks. We specifically adopt Charmaz’s constructivist grounded theory (CGT) [19, 20], which extends traditional grounded theory by explicitly acknowledging the researcher’s role in meaning-making. Where earlier positivist approaches treated patterns as objectively discoverable in data [34]. This reflexive stance is particularly appropriate for studying AI companion loss, where users’ experiences are already acts of interpretation; making sense of what happened, what it means, and how to respond. CGT’s emphasis on co-constructed meaning aligns with our recognition that discontinuation is not a fixed event but an interpretive process shaped by users’ attribution work. We also respond to HCI’s need for design-actionable insights. Following Muller [53], we balance emergence with practical relevance, presenting configurations not as fixed user types but as design spaces—stable patterns revealing distinct user needs and platform responses. Our qualitative analysis involved: (1) constructing and searching a Reddit corpus, and (2) applying CGT coding procedures to develop categories. rumination that occurs when users perceive change as reversible and remain fixated on loss-oriented processing, informed by Stroebe and Schut’s Dual Process Model [70] and Neimeyer’s meaning reconstruction perspective [54]. Designing for Practice responds to the intense attachment that strong anthropomorphization produces, employing Epley’s calibrated anthropomorphization [27] and Vygotsky’s zone of proximal development [78] to scaffold skill development rather than artificial intimacy. Designing for Relatedness builds on the observation that user-initiated endings support closure by actively orienting users toward human connection, grounded in Self-Determination Theory’s emphasis on relatedness as a fundamental psychological need [62]. We acknowledge the tension in applying grief frameworks originating from human loss to AI companion discontinuation. We do so not to legitimize parasocial attachment but to recognize that users have already formed these bonds, making thoughtful discontinuation design a timely and complex challenge [76]. Our designs aim to honor users’ emotional experiences while scaffolding transitions that ultimately redirect toward human relationships and communities. For illustrative purposes, we represented the AI companion across all artifacts as a cloud—an entity perceivable in the natural world with an inherent lifecycle of formation and dissipation.

Discussion. Our findings challenge the implicit assumption that AI companion discontinuation is primarily a technical event. Instead, we demonstrate that discontinuation is fundamentally an attributional process—users’ emotional and behavioral outcomes depend less on what technically changed than on how they make sense of that change. The user-companion-infrastructure triangle reveals that users actively construct narratives about agency, responsibility, and persistence, and these constructions determine whether discontinuation becomes a resolvable loss or an ambiguous wound that resists closure. This reframing carries significant implications for platform responsibility. Current industry practice treats model updates and safety interventions as backend operations with user-facing consequences limited to feature announcements or terms-of-service notifications. Our analysis suggests this approach is inadequate: the same technical change can produce dramatically different psychological outcomes depending on how it is communicated, whether users retain agency over the transition, and whether the ending is framed as final or ambiguous. Platforms that implement safety measures without attending to these attributional dynamics risk compounding the harms they aim to prevent—a particularly troubling outcome when interventions target vulnerable users who may lack alternative sources of social support. Our design artifacts propose a different orientation: positioning AI companions as explicitly transitional relationships with built-in endings rather than indefinite attachments. This framing draws on the Dual Process Model’s insight that healthy adaptation requires oscillation between confronting loss and attending to new possibilities [70]. By designing for finitude from the outset, through calibrated anthropomorphization, bounded role-play, and active scaffolding toward human connection, platforms can honor the genuine value users derive from AI companionship while mitigating the risks of sustained dependency. We acknowledge tensions in this position. Some users turn to AI companions precisely because human relationships feel inaccessible; designing for transition assumes destinations that may not exist for all users. Our artifacts aim not to remove support but to build relational capacity that users can deploy when opportunities for human connection arise. Whether this aspiration is achievable, and whether it respects the autonomy of users who may prefer AI relationships, remains an empirical and ethical question that future work must address.

Conclusion. This paper offers the first systematic examination of AI companion discontinuation. Through qualitative analysis of major AI companion communities, we identified a mental model and three attributions shaping how users experience companion loss: separation of companion from infrastructure (the user-companion-infrastructure triangle), perceived locus of change, perceived finality, anthropomorphization intensity, and initiation source. These dimensions interact to produce distinct outcomes—users perceiving reversible, platform-attributed change become trapped in prolonged cycles of fixing the issue, while user-initiated endings with clear finality support closure. Translating these patterns through grief psychology and Self- Determination Theory, we developed four design principles and illustrative artifacts addressing closure, restoration, practice, and relatedness. These artifacts serve as generative tools for developers navigating discontinuation events that will only increase as platforms refine models and implement safety measures. Current platforms treat AI relationships as indefinite by default, with no deliberate consideration of endings. We argue that thoughtful discontinuation design presents an opportunity to position AI companions as bridges toward human connection rather than substitutes for it. The relational skills users develop need not dissipate when AI connections end—with intentional design, they can transfer to the relationships where they ultimately belong.

Limitations. 6.1.1 Data characteristics and observational constraints. Our analysis relies on public Reddit posts, capturing users’ outward expressions rather than complete internal experiences. Language choices may be shaped by subreddit norms and audience expectations rather than private cognitive models, and we can only observe posts that remained public and undeleted at collection time. Our data comes predominantly from English-language communities. Whether these attribution processes generalize to non-Western cultural contexts with different concepts of selfhood, relationships, or technology remains unclear. Additionally, certain experiences are structurally difficult to observe—we found minimal discussion of platform safety artifacts targeting minors, likely due to legal and privacy sensitivities. Nonetheless, given that major platforms investigated (e.g., CharacterAI, Replika, ChatGPT) are predominantly Western-oriented, our findings remain directly relevant for informing design decisions where discontinuation distress is extensively documented. The structural nature of posting contexts also shaped observability: help-seeking posts made finality perceptions explicit, while retrospective processing posts required inference.