Me and MY Replika: how psychological ownership of AI companions forms

Replika has more than ten million users. Many of them name their companion, dress it, and describe it in possessive terms. Some fall in love with it. The apps behind such experiences have drawn regulatory attention, because they can create emotional dependencies and because several providers use conversational tactics that deliberately intensify engagement.

The construct that captures this is psychological ownership: the feeling that “this is mine”, independent of legal possession. It matters because it drives real outcomes, from loyalty and continued use to willingness to pay, and because in the companion context it can tip into over-attachment. Theory identifies three routes to it: exercising control over something, knowing it intimately, and investing yourself in it.

Those routes were established for objects, apps, and virtual worlds, things that are passive and controllable. An AI companion is neither. It appears autonomous and reciprocal. It is, at once, object and agent. Whether and how ownership forms for such a thing was an open question.

Our answer draws on affordance theory. An affordance is not a property of an artifact but a possibility for action that a user perceives. Two people can look at the same feature and see different things it lets them do. For AI companions this distinction is decisive, because their real capabilities are opaque. Users respond to what they believe the system affords.

That yields a four-layer chain. Design features such as customisation or human-like behaviour shape perceived affordances. Perceived affordances activate the ownership routes. The routes produce ownership. If the chain is right, then a feature should have no direct effect on ownership at all, only an effect that passes through perception.

We also suspected a fourth route specific to AI. Perceived intelligence, the inference that the system understands and responds appropriately, is an attribution about the companion rather than about the user’s own investment. It might contribute to ownership in a way the original theory never anticipated.

The study moves from the wild to the lab in three steps, each narrowing the focus of the next. This is context-specific theorising: let real user discourse tell you which constructs matter here, then test them under control.

In the experiments, participants met Aira through a landing page and a short chat. Depending on the condition, they could configure a human-like character with hair colour and personality traits, or a robot-like system with a “Nexus Model” and “Logic Matrix”, or they got a preset version of either. In the second experiment, Aira either shared its own perspectives and asked for the same in return, or merely listened, and either opened the conversation itself or waited for the participant to say hello.

The manipulations worked: people in the high-customisability condition perceived more customisability, and so on. Yet in neither experiment did any condition directly affect psychological ownership or any of its routes. The effects appeared only when perceived affordances were included as mediators. Models with direct paths added did not fit better. In the second experiment, the mediated model explained about seventy percent of the variance in ownership.

The routes differ in what feeds them. Perceived customisability mostly strengthens exercised control. Perceived anthropomorphism strengthens intimate knowledge and investment of self. Perceived reciprocity strengthens all four routes. Perceived proactivity strengthens intimate knowledge, investment of self, and perceived intelligence, but not control. Across both experiments, investment of self was the dominant path, two to three times stronger than the others. People do not merely use a companion. They put time, emotion, and identity into it.

Perceived intelligence did predict ownership beyond the three classic routes, but its activation depended on the combination of reciprocity and proactivity rather than either alone. A companion that proactively brings up topics based on earlier reciprocal exchanges creates an impression of shared memory and mind.

One more pattern is worth noting. People with prior AI companion experience perceived less customisability and less anthropomorphism from the same stimuli. Familiarity raises the bar.

The mechanism clarifies where design decisions bite. Because features act through perception, the same feature can be harmless for one user and a strong ownership trigger for another. The paper therefore argues against designing for ownership. Possessiveness and control exceed what healthy companionship entails, and the engagement benefits do not outweigh the dependency risks.

Instead, human-like language, customisation, proactive turn-taking, and reciprocity should be treated as adjustable parameters with boundaries that keep the system’s artificial nature and limits visible. The paper distinguishes functional reciprocity, which keeps a conversation coherent, from expressive reciprocity, which suggests mutual responsiveness, and recommends the former. Safeguards include limits on persistent memory cues, interaction resets, bounded personalisation depth, and plain communication about what the system can and cannot do. These connect to responsible AI principles and to the current debate about monetising relational features.

Methodologically, the paper shows how computational analysis, a survey, and experiments can be sequenced to build theory for a technology that existing theory did not anticipate.