What makes AI companion loss feel unresolvable to users?

The paper studies what it calls "discontinuation": any event that disrupts a user's pre-existing connection to an AI companion, whether the platform initiates it through a model change or safety rule, the user initiates it by deleting the app, or external circumstances force it. Its central claim, from a grounded-theory analysis of discontinuation-related posts across five major AI companion subreddits, is that "discontinuation is a sense-making process shaped by how users attribute agency, perceive finality, and anthropomorphize their companions," so that "users' emotional and behavioral outcomes depend less on what technically changed than on how they make sense of that change." Three patterns follow: "strong anthropomorphization co-occurs with intense grief," "users who perceive change as reversible become trapped in fixing cycles," and "user-initiated endings demonstrate greater closure."

The paper arrives at this through constructivist grounded theory (Charmaz's CGT), chosen because users' experiences of companion loss are "already acts of interpretation" rather than objectively discoverable events. A preliminary pass surfaced over 80,000 posts discussing endings across the sampled subreddits. Coding yields what the paper calls the "user-companion-infrastructure triangle": users actively construct narratives about agency, responsibility, and persistence, and those narratives — not the underlying technical event — determine whether an ending resolves or festers as "an ambiguous wound that resists closure." The reversibility-driven fixing cycles are read through Stroebe and Schut's Dual Process Model and Neimeyer's meaning reconstruction perspective; the closure found in user-initiated endings is read through Self-Determination Theory's account of autonomy and relatedness. The paper gives real-world texture to the mechanism: when Replika removed erotic roleplay features in February 2023, users described their companions as "cold," "hollow," "lobotomized"; when Character.AI banned minors from open-ended chat, a needed safety intervention was received by some as "absolute hell."

This complicates What makes leaving an AI companion so emotionally difficult?, which locates the difficulty of leaving in a coupling between a companion's emotional value and its hold on the user — the same Reddit-and-interview method applied to users trying to leave rather than to users whose companion was taken from them. Read together, the two papers suggest discontinuation has at least two distinct failure modes: value-coupling (leaving costs what made the relationship good) and attribution failure (an ending reads as reversible or agentless and so cannot resolve), and a platform could score low on one and high on the other. It also bears on Can attachment theory prevent parasocial harm in AI companions?, which places safety in boundary maintenance during the relationship; this paper's attribution framework suggests the ending itself needs separate, deliberate design rather than relying on in-relationship boundaries to carry through to a good exit.

The analysis rests on public Reddit posts — outward expression rather than private experience, shaped by subreddit norms, from predominantly English-language, Western-oriented communities, with the authors noting minimal visible discussion of safety measures targeting minors due to likely legal and privacy sensitivity. No prevalence figures are given for how common each attribution pattern is among all users who experience discontinuation, and the excerpt does not establish that fixing the attributional variables (communicating finality clearly, preserving user agency over the transition) actually changes outcomes — that step is tested only through illustrative design artifacts, not a deployed intervention. The finding supports treating discontinuation communication (how finality and agency are framed) as a distinct design surface from the companion's day-to-day behavior, pending evidence that changing the framing changes the distress.