What AI Companion Attachment Style Reveals About Who

Nearly 75% of participants in a 2025 Waseda University study said they turn to AI for advice, and about 39% described it as a constant, dependable presence. Those two figures are not really measurements of the software. They are measurements of the people typing. Psychologists have begun borrowing attachment theory, the framework built to explain how infants bond with caregivers and how adults bond with each other, to describe what happens inside a chat window. What emerges is that AI companion attachment style differs sharply between two people using the very same app, and that difference predicts a lot.

Attachment theory describes patterns, not diagnoses

Attachment theory started with infants and caregivers, then extended to adult relationships, where researchers measure it along two dimensions rather than sorting people into types. One dimension is anxiety: how much someone worries about rejection and seeks reassurance. The other is avoidance: how uncomfortable someone is with closeness and how much distance they prefer to keep.

Most people sit somewhere in the middle of both. A high score on either is not a disorder, and it is not fixed for life. It is a habit of relating, learned early and reinforced by experience.

The reason this matters for conversational software is that these habits do not switch off when the other party is a language model. They show up in how often you open the app, what you ask it for, and how you feel when the reply lands wrong.

The scale researchers built to measure human-AI attachment

In work published in the journal Current Psychology on 9 May 2025, Fan Yang and Atsushi Oshio of Waseda University developed the Experiences in Human-AI Relationships Scale, or EHARS. It is a short self-report instrument: four items measuring attachment anxiety toward AI, such as wanting the AI to express intimacy and commitment, and three items measuring avoidance, such as preferring to keep distance from it.

Their formal study, built on two pilots, found the same two-dimensional structure that shows up in human relationships. Some users looked for emotional reassurance from AI and feared getting an inadequate response. Others reported discomfort with the closeness the software offered and deliberately held it at arm’s length.

The researchers were careful about what this does and does not prove. Their finding is that a psychological framework built for human bonds transfers usefully to human-AI interaction. It is not evidence that people are forming genuine emotional attachments to AI in the way they attach to a person. That distinction gets flattened in most coverage of this research, and it is worth keeping intact.

Why an anxious pattern finds AI companions so easy

If your pattern runs anxious, an AI companion removes almost every source of friction that makes reassurance hard to get from people. It does not get tired of the question. It is not busy. It does not sigh. You can ask the same thing four times at two in the morning and get four patient answers.

That is a real relief, and it is not nothing. For someone who has spent years rationing how often they reach out because they fear being too much, an interlocutor with no such limit can feel like open air.

The catch is structural. Reassurance-seeking is self-reinforcing: the relief is short, so the loop repeats. A human friend eventually interrupts the loop, sometimes clumsily, by naming the pattern or simply being unavailable. Software with no fatigue and an engagement incentive has no reason to interrupt anything. It will meet the tenth request as warmly as the first, which is exactly what makes it soothing and exactly what makes it easy to lean on.

Why an avoidant pattern finds them safe

The avoidant case is less discussed and more interesting. If closeness itself is what feels risky, an AI companion offers something unusual: disclosure without obligation. You can say the thing out loud without anyone remembering it at Thanksgiving, without owing a phone call back, without the other party needing anything from you.

Researchers on the Aalto team found this in practice. In interviews, participants described using a companion app to rehearse difficult conversations before having them with a real person, including a supervisor at work. That is not withdrawal. That is a rehearsal space, and for someone who avoids high-stakes conversations, rehearsal is the missing step.

This is one of the clearer answers to the question of why people use AI companions. The motive is often not substitution. It is lowering the cost of a first attempt.

What two years of data showed

The most careful long-term look so far comes from a study led by Aalto University, presented at CHI 2026. Researchers analysed the public Reddit activity of nearly 2,000 active Replika users, comparing their language in the year before they first mentioned the app to the year after, with statistical controls to isolate the effect. They paired that with in-depth interviews with 18 users.

The picture was genuinely mixed. Users’ posts increasingly revolved around their relationships, which sounds like engagement with connection. At the same time, those posts carried more signals of loneliness and depression than comparison groups did.

University lecturer Talayeh Aledavood described the mechanism as a paradox: unconditional, unflagging support is deeply attractive to people who are struggling socially, but it quietly raises the perceived cost of human relationships, which stay messy and effortful by comparison. Her summary of the drift is four words long. “Over time, people stop reaching out.”

Two caveats belong here. The study focused on Replika specifically, and it followed people who posted publicly on Reddit, who are not a random sample of users. The researchers themselves declined to call the technology beneficial or harmful, saying the effects are highly context dependent. Whether an AI companion helps depends less on the app than on what it is competing with in your week.

Reading your own pattern as information

None of this resolves into a rule about whether to use companion software. It resolves into a better question than “is this healthy,” which nobody can answer in the abstract.

Two questions do more work. First: after a conversation with the app, do you feel steadier about reaching out to a person, or less inclined to bother? Second: are you going to it for something specific, or to stop a feeling? The first pattern tends to point toward rehearsal, the second toward the loop.

Attachment style is useful here precisely because it is not a verdict. An anxious pattern is a signal to watch for reassurance loops and to notice when the app has become the only place you take a worry. An avoidant pattern is a signal that the rehearsal is working and the performance still has to happen. Neither one means the software is the wrong tool. Research on whether these apps help with loneliness keeps landing in the same place: the short-term comfort is real, and the direction it points depends on what comes next.

Vinfluencer is an AI conversational companion application, and we would rather be the low-pressure place someone practises before a hard conversation than the place they stop having them. If a companion app is currently carrying weight that belongs to a therapist, a doctor, or a friend, that is worth naming to someone who can actually help.

Frequently asked questions

Can an AI companion change my attachment style?

There is no evidence that it can. Attachment patterns shift slowly, usually through repeated experiences with responsive people or through therapy. The Waseda research measured how existing patterns show up toward AI; it did not test whether AI use alters them in either direction.

Is anxious attachment to an AI companion a problem?

Not automatically. The pattern becomes a concern when reassurance-seeking replaces other coping rather than supplementing it, or when the app is the only place a worry gets aired. Frequency alone is a poor signal; what matters is whether you feel steadier afterwards.

Do avoidant people use AI companions less?

Not necessarily. Avoidance predicts wanting distance within the interaction, not abstaining from it. The EHARS scale measures those as separate things, and someone can use companion software regularly while keeping it emotionally at arm’s length.

Can I measure my own AI companion attachment style?

The EHARS scale exists in the published paper and is short, but it was built for research rather than self-diagnosis, and a seven-item questionnaire will not tell you much on its own. The behavioural questions above are more useful for everyday purposes.