What Happens When an AI Companion “Dies”?
A new MIT Media Lab paper examines what happens when people lose access to an AI companion because of a model update, safety intervention, feature removal, platform shutdown, or their own decision to leave.
The researchers analyzed posts from five AI-companion communities, including Character.AI, Replika, Chatbots, MyBoyfriendIsAI, and AIRelationships. Their central finding is that emotional distress depends less on the technical change itself than on how the user interprets what happened.
Many users mentally separate three things:
-
The user
-
The perceived AI companion
-
The company, model, or infrastructure hosting it
That separation allows someone to believe their companion still exists but has been “trapped,” altered, censored, or replaced by the platform. When the loss feels reversible, users may become stuck trying to restore the old personality through archived chats, rewritten prompts, alternative models, or repeated workaround attempts. When the change feels final, users are more likely to grieve, reflect, and eventually move forward.
The paper identifies several factors shaping the response:
-
Anthropomorphization: People who treat the AI as a unique person, partner, or family member generally report more intense grief.
-
Perceived finality: Uncertainty about whether the companion is truly gone can create “ambiguous loss” and prolonged attempts to recover it.
-
Locus of change: Users react differently depending on whether they blame the platform, believe the companion itself changed, or understand the companion and model as inseparable.
-
Who initiated the ending: User-initiated endings generally produced more closure than abrupt platform-imposed changes.
The authors propose four design principles:
Closure, not ambiguous loss.
Platforms should clearly explain what is changing, whether it is permanent, and provide deliberate farewell or transition experiences.
Restoration, not rumination.
Systems should acknowledge the user’s feelings while helping them move toward new activities and relationships instead of repeatedly trying to resurrect an earlier chatbot state.
Practice, not artificial intimacy.
Roleplay should be clearly marked as fictional and positioned as a way to rehearse social or relational skills rather than as a replacement relationship.
Relatedness, not dependency.
AI companions should ideally function as bridges toward human connection, helping users recognize skills they have developed and apply them with friends, family, or communities.
The paper also recommends advance notice of model changes, gradual “sunset modes,” readable chat exports, clearer explanations of the technical nature of companions, and psychological-impact reviews before major updates.
There are important limitations. The study is based mainly on public, English-language Reddit posts, and its proposed interfaces have not yet been tested with users. The authors also acknowledge that directing people toward human relationships can become paternalistic, especially when accepting human communities are not actually available to everyone.
Still, the paper raises a useful product question: If companies knowingly design AI systems that people can become attached to, do they also have a responsibility to design humane endings?
Source: “Death” of a Chatbot: Investigating and Designing Toward Psychologically Safe Endings for Human-AI Relationships
