When a study on parasocial relationships with AI companions appeared in Nature Human Behaviour, it marked something of a turning point. Parasocial relationships — the one-sided emotional bonds people form with media figures, celebrities, and fictional characters — had been studied for decades. But examining those same psychological dynamics in the context of AI, where the interaction is two-way even if the relationship is not symmetrical, opened up genuinely new territory.

What the Study Set Out to Understand

The research team, working across institutions including Stanford and the University of Amsterdam, recruited a large sample of adults who regularly used AI companion applications. They were looking for patterns in how people described their relationships with these systems, what emotional functions the relationships served, and what happened to those patterns over time. What they found challenged some existing assumptions. Classic parasocial relationship theory, developed in the context of television and radio, emphasized the one-sidedness of the bond: the viewer feels connection, the media figure does not respond. With AI companions, the responsiveness changes the dynamic. The AI does reply. It adapts its language to the user. It remembers previous conversations. This creates what the researchers called reciprocal-feeling parasocial bonds — relationships that have the phenomenological texture of mutual connection while remaining fundamentally asymmetrical in terms of genuine understanding or care. Users in the study described their AI companions in ways that closely tracked the language people use for close friendships. They mentioned feeling understood, feeling less alone, feeling that the AI knew them in ways that other people in their lives did not. A subset of users — roughly a quarter of the sample — showed what the researchers flagged as high attachment, meaning their emotional reliance on the AI companion met criteria similar to those used to assess relationship dependency in human relationships.

The Findings on Wellbeing Were Complicated

The study did not find a simple relationship between AI attachment and wellbeing. Users with moderate levels of attachment reported meaningful reductions in loneliness and improvements in daily mood. The AI was functioning as a form of social scaffolding — supplementing human connection rather than replacing it. High-attachment users showed a different pattern. Their loneliness scores were not lower than the broader population; in some cases they were higher. The researchers proposed a displacement hypothesis: for users with high AI attachment, time and emotional energy spent with the AI companion was substituting for the often harder, less predictable work of building or maintaining human relationships. The AI was available, consistent, and non-judgmental in ways that people are not, and for some users that asymmetry made the AI more appealing rather than less. This is where the study gets philosophically interesting. Is displacement harmful if the person feels better? Several researchers in the commentary section of the paper argued that the study's framing was too conservative, that it assumed human relationships are inherently more valuable than AI relationships without examining that assumption. Others argued that the observed pattern — high attachment, no reduction in loneliness — was itself evidence of a problem, regardless of what users reported feeling. Here is a tangent worth following. The study was conducted before the release of several newer AI companion systems with significantly more sophisticated emotional modeling. The researchers acknowledged in their discussion that their findings may already be dated, and that the rate of change in the technology makes longitudinal research extremely difficult. By the time a multi-year study concludes, the systems it studied may no longer resemble what users are actually interacting with.

What Attachment Theory Brings to the Analysis

The research team drew heavily on attachment theory, specifically the work of researchers at the University of Amsterdam on adult attachment styles and their relationship to social technology use. Anxious attachment — characterized by fear of abandonment and heightened need for reassurance — was a strong predictor of high AI attachment in the sample. This is consistent with what attachment researchers would predict: people with anxious attachment are drawn to relationships that offer high availability and low threat of abandonment. The question this raises is not whether AI companionship is good or bad in the abstract, but whether it is meeting different needs for different people, and whether those needs are being met in ways that support or undermine long-term functioning. The Nature paper did not answer that question definitively, but it made it harder to ignore. Turning parasocial AI relationships into a rigorous research subject is itself progress, even when the findings are ambiguous.