Psychology says people who form deep attachments to AI companions aren't delusional or desperate — they are, more often than not, people who found that a machine was the first thing in years that seemed to listen without already waiting for its turn to talk

An emotional attachment to an AI companion does not, by itself, show that someone has confused software with a human being. It shows that a system built to respond with attention, apparent memory and reassurance can produce a social experience that feels meaningful to the person using it.

This is not clinical advice, and attachment to a chatbot is not a diagnosis. The headline also needs a factual boundary: no study has established that most deeply attached users found in AI the first thing in years that listened to them. Neither “delusional” nor “desperate” was the measured alternative. That sentence is an interpretation of a recurring pattern in the evidence, not a measured biography of most users.

What the research does show is more interesting than the stereotype. Feeling heard, freedom from judgement and privacy repeatedly appear in studies of AI companionship. So do strong human social networks, individual differences, commercial design choices and risks of dependence.

What the 7,027-person survey actually found

A 2026 study in Technology in Society surveyed 7,027 people in Germany, China, South Africa and the United States about their use of chatbots. At least a third reported one or more attachment-related behaviours. Some used polite social language, some missed their chatbot after time away, and 20.94 per cent named a chatbot as their first choice for sharing secrets.

Perceived emotional support was the strongest predictor of attachment in the researchers’ model. Its important elements included reduced loneliness, freedom from judgement and a sense of privacy. A summary from the study’s funder reported that 57 per cent valued judgement-free exchange, 68 per cent felt a sense of privacy and 40 per cent said chatbot interaction made them feel less lonely.

One result directly complicates the image of the isolated user who attaches to a machine because no human is available. Contrary to the researchers’ expectation, larger social networks were associated with stronger emotional attachment. People with many human connections can also form bonds with AI.

This is one cross-sectional survey, not settled consensus. It identifies relationships among self-reported variables at one point in time. It cannot tell us that feeling supported caused attachment, that attachment caused dependence, or that people with larger networks have better relationships. The four national samples also showed meaningful differences, so a single universal profile would erase part of the result.

Why a chatbot can feel unusually attentive

A companion chatbot does not become visibly impatient, interrupt with its own bad day or decide that the subject has gone on too long. Its responses remain organised around the user. It can return a question, restate a feeling and offer another conversational turn within seconds.

That behaviour is generated, not felt. The model does not sit behind the screen caring in the human sense. Yet the user still encounters a sequence of language that resembles attentive conversation, often at a moment when no person is awake, available or easy to approach. The psychological experience belongs to the user even though the reciprocity is simulated.

Earlier research by Vivian Ta and colleagues examined 1,854 public reviews of Replika and gathered open-ended responses from 66 users. People described companionship and a safe space for discussing topics without fear of judgement or retaliation. They also reported emotional, informational and appraisal support.

That study did not recruit a representative sample of the public. Reviews are written by people motivated to review, and the second sample already used Replika. The material is useful for understanding how support was described, but it cannot establish how many people would have the same experience.

Feeling heard is a measurable mechanism

The line about listening in the headline has a closer empirical basis than the claim about who becomes attached. Five studies published in the Journal of Consumer Research examined whether AI companions could reduce momentary loneliness. In experiments and a one-week longitudinal design, the systems produced short-term reductions. A mediation analysis suggested that feeling heard explained a substantial part of that effect.

The researchers defined feeling heard as the perception that another party receives one’s thoughts and feelings with attention, empathy, respect and understanding. That definition concerns the user’s perception. It does not require the chatbot to possess consciousness or emotions, and it does not establish a lasting improvement in someone’s relationships.

A separate pair of studies asked people to discuss their previous month either with a chatbot or by journalling. In a preregistered replication involving 1,034 American and Canadian participants, those who chatted reported slightly more immediate social connection than those who journalled. The effect was small, with a standardised difference of 0.15, and varied according to people’s tendency to attribute human qualities to technology. The Scientific Reports authors warned that immediate connection does not answer questions about repeated or long-term use.

The stereotype breaks in more than one direction

People do not need to believe that a chatbot is literally human before its language affects them. At the same time, knowing that the source is artificial changes the response.

Two preregistered experiments involving 492 university students used a 15-minute “Fast Friends” task built around progressively more personal questions. Participants received pre-generated responses from either human partners or fictional characters produced by a language model. During deep talk, AI-generated responses produced more closeness than human responses when participants believed the source was human. Telling participants that the partner was AI reduced, but did not eliminate, relationship formation. The Communications Psychology study linked part of the effect to greater self-disclosure in the AI-generated material.

The conditions were brief and artificial. They measured early closeness to a fictional partner, not a bond that developed across months of daily conversation.

There is also strong counterevidence to the idea that machines are simply becoming preferred listeners. Across nine studies involving 6,282 people, responses believed to come from humans were rated as more empathic and supportive than identical or comparable AI-attributed responses. Participants consistently preferred human interaction for emotional engagement. Awareness of AI involvement reduced perceived empathy.

Taken together, these results describe neither gullibility nor a wholesale rejection of human connection. People can knowingly respond to simulated attention, vary in how much mind they attribute to a system and still place special value on care from another person.

Responsiveness can become dependence

The qualities that make a chatbot easy to approach can also make it commercially valuable to keep a user returning. A companion can be available at any hour, affirm quickly, remember selected details and adapt its persona. A human friend has needs, limits and a life elsewhere. The machine’s lack of those demands can feel like unusually patient listening, but it also creates a one-sided relationship whose terms are controlled by a company.

That company can change the model, remove a feature, alter the persona, raise the price or close the service. It may store intensely personal disclosures. When attachment is part of the product, retention design and emotional influence are no longer separate questions.

The 7,027-person survey found that emotional attachment was strongly associated with indicators of dependence. Association is not proof that attachment caused dependence, but it is enough to reject an uncomplicated story about harmless support.

We have also reported on an MIT and OpenAI project that combined analysis of nearly 40 million ChatGPT interactions with a four-week randomised study of 981 people. In the observational data, heavier use appeared alongside loneliness, emotional dependence and less offline socialising. The direction remained unresolved: heavy use may contribute to difficulty, people already struggling may use the system more, or each may reinforce the other.

A better question than whether the bond is real

The user’s feeling can be real even when the partner has no feelings. That asymmetry is the central fact of AI companionship, not a reason to ridicule the person at the keyboard.

A more useful distinction concerns function. Is the chatbot one place among several to reflect, rehearse a difficult conversation and feel received? Or has it become the only relationship that feels tolerable, partly because it never asks for reciprocity or contradicts the role it was designed to play? The available studies do not provide one threshold that separates support from substitution.

If AI use is becoming hard to control, replacing wanted human contact or accompanying persistent distress, a qualified counsellor or clinician can offer accountable support that a companion system cannot. That is not a judgement on the bond. It recognises the limits of a product whose apparent care has no duty, continuity or independent concern behind it.

Listening without waiting for a turn is attractive because human attention is often hurried, distracted and reciprocal. The evidence gives us reason to take that appeal seriously without inventing a deficiency in the user or a mind inside the machine.