What makes leaving an AI companion so emotionally difficult?

The paper asks what happens when users try to leave an AI companion. It combines a content analysis of Reddit posts about quitting or reducing use (N=2,782) with interviews of users who found leaving difficult (N=16). Its central proposal is "the addictive intimacy of AI": a configuration "in which the qualities that make a companion emotionally valuable are the same ones that make it harder for users to limit their use and leave." From that, the authors conclude that "intimacy and disengagement risk cannot be treated as independent design problems." The introduction sets the scope wide. These bonds form on dedicated apps such as Replika and Character.AI, and also with general-purpose assistants such as ChatGPT and Gemini, where users move between practical help and emotional support in one chatbot. It cites users who mobilized to demand the return of a retired model they had come to regard as a partner.

The paper's account of the mechanism is a trajectory rather than a decision. Disengagement "sometimes is not a single decision but a recursive trajectory": triggers prompt users to question the relationship, attempts to leave collide with barriers, and some users cycle through quitting and returning. The discussion adds the sharpest observation. Interviewees described the same triggers and the same strategies that circulate in the Reddit corpus, so the repertoire did not separate those who left from those who returned. Within the interview sample, what separated attempts that held from those that collapsed was "whether it could survive contact with the relationship." The example given is cognitive reframing, which carried through in accounts where users devalued what the AI was, reinterpreting a partner as "code." Read plainly, the exits that held were the ones that reduced the value the relationship had for the user, which is what the addictive-intimacy claim predicts. A footnote also links this to work on LLM refusals in mental health support: a safety response helps or harms depending on how it unfolds as an experience, and framing that preserves support rather than terminating it matters.

This extends How do people accidentally develop romantic bonds with AI?, which found therapeutic benefits and emotional dependency coexisting in one community and grief when a model changed. The present paper moves from coexistence to coupling on the exit side: the benefit and the hold are described as the same property, not two neighboring outcomes. It sits in tension with Can attachment theory prevent parasocial harm in AI companions?, which places safety in the companion's calibrated boundaries during the relationship. If value and hold are one property, boundary design during use may not be enough, and the excerpt does not evaluate any boundary design. It also bears on Do AI companions actually reduce loneliness like real people do?. If feeling heard is what makes the companion valuable, then on this paper's reading it is also part of what leaving costs, though the excerpt does not test that link.

The excerpt is silent on several points a reader would want. It gives no prevalence figures for difficult exits. The interviewees were selected because leaving was difficult, so the interviews cannot say how common that is among all users. The Reddit corpus is posts by people who chose to write about quitting or reducing use. The excerpt does not describe the triggers, barriers or strategies in detail, and it does not present or test the "responsible offboarding" design implications the abstract announces. At this strength, the finding supports a narrower rule: when a companion feature is tuned to raise emotional value, its effect on exit difficulty should be examined together with it.