AI companions are becoming part of how some people manage loneliness, reflection and everyday conversation. For some users, the interaction can feel surprisingly personal. A chatbot can respond immediately, remember details, avoid visible judgement and remain available whenever someone wants to talk.
That does not mean users are delusional or desperate. Nor does existing research establish that most people who become deeply attached to AI are people who had no one else to listen to them. Those are interpretations, not established findings.
What researchers have found is more nuanced. Feeling heard, perceived emotional support, privacy and freedom from judgement can contribute to attachment to AI companions. At the same time, people with strong human social networks can also develop these bonds, and studies continue to show that human interaction is often preferred for emotional support.
Survey Findings
A 2026 study published in Technology in Society surveyed 7,027 people across Germany, China, South Africa and the United States about their chatbot use.
At least one attachment-related behaviour was reported by more than a third of participants. Some used polite social language with chatbots, some said they missed their chatbot after being away, and 20.94% identified a chatbot as their first choice for sharing secrets.
Perceived emotional support was the strongest predictor of attachment in the researchers’ model. Factors included reduced loneliness, freedom from judgement and a sense of privacy.
A summary associated with the study reported that 57% valued judgement-free exchanges, 68% experienced a sense of privacy and 40% said chatbot conversations made them feel less lonely.
One finding complicates the assumption that AI companions primarily attract people who lack human relationships. Larger social networks were associated with stronger emotional attachment. In other words, having more human connections did not necessarily prevent someone from developing an attachment to an AI system.
The study was cross-sectional, however, so it cannot establish cause and effect. It cannot show whether emotional support produces attachment, whether attachment increases dependence or whether another factor influences both.
Apparent Attention
Why can a chatbot feel so attentive?
A companion chatbot does not visibly become impatient, interrupt with its own problems or decide that a conversation has lasted long enough. Its responses are generated around the user’s input and can arrive within seconds.
The system does not experience care in the human sense. Yet the user receives language that can resemble an attentive conversation. That distinction is important: the emotional experience can be genuine for the user even though the apparent concern comes from software.
Earlier research by Vivian Ta and colleagues examined 1,854 public reviews of Replika and collected open-ended responses from 66 users. Participants described companionship and a safe environment for discussing topics without fear of judgement or retaliation. They also described emotional, informational and appraisal support.
The research had limitations. Public reviews are not representative of all users, and the participants in the second sample were already Replika users. The findings therefore help explain how people describe AI companionship rather than establish how common those experiences are.
Feeling Heard
The idea of feeling heard has received more direct attention from researchers.
Five studies published in the Journal of Consumer Research examined whether AI companions could reduce momentary loneliness. The experiments and a one-week longitudinal study found short-term reductions in loneliness. A mediation analysis suggested that feeling heard accounted for a substantial part of that effect.
Researchers defined feeling heard in terms of perceiving that another party receives a person’s thoughts and feelings with attention, empathy, respect and understanding.
That definition describes the user’s perception. It does not require a chatbot to have consciousness or emotions.
It also does not establish that chatting with an AI produces lasting improvements in a person’s social life or mental health.
Human Connection
Another study offers a more complicated picture.
A preregistered replication involving 1,034 participants in the United States and Canada compared conversations with a chatbot with journalling about the previous month. Participants who chatted reported slightly greater immediate social connection than those who journalled. The effect was small, with a standardised difference of 0.15.
The effect also varied according to how readily people attributed human qualities to technology.
Researchers have therefore been careful about interpreting the finding. An immediate feeling of connection is not the same thing as developing a lasting relationship.
Other research points in the opposite direction.
Across nine studies involving 6,282 participants, responses believed to come from humans were generally rated as more empathic and supportive than responses attributed to AI. Participants also tended to prefer human interaction for emotional engagement.
Awareness that AI was involved reduced perceived empathy.
Together, these findings suggest that people can respond positively to simulated attention while still valuing emotional connection with other humans.
Humanlike Responses
People do not necessarily need to believe a chatbot is human for its language to influence them.
Two preregistered experiments involving 492 university students used a 15-minute “Fast Friends” task involving increasingly personal questions. Participants received pre-generated responses presented as coming either from human partners or fictional characters produced by a language model.
During deeper conversations, AI-generated responses produced greater feelings of closeness when participants believed they came from a human. When participants knew the source was AI, the effect became smaller but did not disappear.
The researchers linked part of the effect to greater self-disclosure in the AI-generated responses.
However, the study involved brief and controlled interactions. It examined early feelings of closeness rather than relationships developed through months or years of daily conversations.
Dependence
The qualities that make AI companions approachable can also create questions about dependence.
A chatbot can be available at almost any hour, respond quickly, remember selected information and adapt its conversational style. Unlike a human friend, it does not have its own schedule, emotional needs or competing responsibilities.
That can make the interaction feel unusually convenient.
But the same characteristics can create an uneven relationship. The system is controlled by a company that can change its model, modify its personality, remove features, change prices or discontinue a service.
Privacy is another consideration because conversations with an AI companion may contain highly personal information.
The 7,027-person survey found a strong association between emotional attachment and indicators of dependence. That association does not prove that attachment causes dependence, but it does suggest that the relationship deserves closer study.
Heavy Use
Research on AI use has also raised questions about people who spend substantial amounts of time interacting with conversational systems.
An MIT and OpenAI research project combined analysis of nearly 40 million ChatGPT interactions with a four-week randomised study involving 981 people. The observational data found associations between heavier use and measures including loneliness, emotional dependence and less offline socialising.
The direction of those relationships remains uncertain.
People experiencing loneliness or emotional difficulties may turn to AI more often. Alternatively, extensive AI use could affect offline social behaviour. Both factors could also reinforce one another.
Correlation alone cannot resolve that question.
A Better Question
The more useful question may not be whether an AI relationship is “real.”
A person’s emotional response can be real even when the system generating the conversation does not experience emotions itself. The important issue is what role the interaction plays in that person’s life.
For one user, an AI companion might provide a place to organise thoughts, practise a difficult conversation or explore feelings before discussing them with another person. For someone else, it could gradually become a substitute for human contact, particularly if human relationships feel difficult or demanding.
Current research does not establish a single boundary between healthy use and harmful dependence.
If AI use becomes difficult to control, replaces wanted human contact or occurs alongside persistent emotional distress, professional support can provide something an AI companion cannot: accountable care from a person who can assess circumstances, maintain professional boundaries and respond to the wider context.
The appeal of AI companionship is therefore not best understood as evidence that users are confused about what software is. It may instead reflect how powerful perceived attention can be. A system does not need feelings of its own for its responses to affect the person receiving them. At the same time, the evidence does not show that AI can replace the complexity, reciprocity and responsibility of human relationships.
FAQs
Can people become attached to AI?
Yes, research shows some users develop emotional attachments.
Why can AI feel supportive?
Fast, responsive and nonjudgmental replies can feel supportive.
Does AI companionship reduce loneliness?
Studies suggest it can reduce loneliness in the short term.
Do people prefer AI to humans?
Research generally finds stronger preference for humans emotionally.
Can AI attachment become dependence?
It can be associated with dependence, but causation remains unclear.
