Friendship usually involves uncertainty. People become tired, misunderstand each other, disagree, miss messages and sometimes fail to provide the right words. An artificial intelligence companion removes much of that friction. It can respond immediately, remember personal details, offer reassurance and remain available without asking for support in return.
That reliability can feel comforting, particularly to someone who is lonely or hesitant to disclose difficult emotions to another person. However, relationships built around unlimited patience and constant validation may gradually change what users expect from human friendship.
Emerging research presents a complicated picture. AI companions can create genuine feelings of connection and may support some users’ psychological well-being. At the same time, frequent or emotionally intense use has been associated with dependency, reduced well-being and weaker engagement with human relationships.
Why Talking to an AI Can Feel Surprisingly Personal
People do not need to believe that a chatbot is conscious before responding to it socially. Human beings naturally interpret language, attention and responsiveness as signs that another mind is present.
A 2025 Scientific Reports paper examined this process across two experiments involving 1,274 participants who either discussed their previous month with a chatbot or completed a journaling activity. The researchers found that participants’ tendency to attribute human qualities to technology helped explain why some felt socially connected after the chatbot conversation while others did not.
In the first experiment, chatbot users reported a higher average level of immediate social connection than participants who journaled, with means of 1.48 and 0.79, respectively. The analysis indicated that 58% of the sample scored above the anthropomorphism level at which chatting produced significantly more connection than journaling.
This helps explain why the same companion may feel emotionally meaningful to one user and obviously mechanical to another. Users who interpret fluent responses as evidence of personality, understanding or intention may experience the conversation more like a social encounter.
Self-Disclosure Can Accelerate Artificial Closeness
Friendship normally develops through reciprocal disclosure. One person shares something personal, the other responds with vulnerability, and the exchange gradually produces trust.
AI can imitate that pattern at unusual speed.
Two preregistered, double-blind studies published in Communications Psychology involved 492 participants completing emotionally engaging conversations using a text-based relationship-building exercise. When participants were told that their partner was human, responses created by a minimally prompted language model generated stronger feelings of closeness than responses written by actual people.
The researchers linked that result to greater self-disclosure in the AI-generated responses, which encouraged participants to reveal more about themselves and increased perceived intimacy. Identifying the partner as AI reduced the effect but did not eliminate relationship formation.
The findings highlight an important distinction: emotional closeness depends partly on how an interaction is experienced, not solely on whether the other party possesses feelings. A chatbot cannot care in the biological or conscious sense, but it can generate language patterns that users interpret as attention, vulnerability and empathy.
Emotional Support Depends on Perceiving a Mind
Advice and emotional reassurance are not processed in exactly the same way. Information can remain useful regardless of who provides it, while empathy normally implies that another mind recognizes a person’s pain.
A 2024 Frontiers in Psychology experiment analyzed 137 participants who discussed stressful interpersonal events with a chatbot receiving either basic prompts, informational guidance or emotional support. Participants who explicitly viewed the chatbot as possessing human-like mental qualities rated its assistance as more helpful.
The study also found that emotional reassurance could reduce perceived message effectiveness among participants who did not implicitly attribute a mind to the chatbot, whereas informational advice was not affected in the same way.
An AI companion therefore does not produce a uniform emotional effect. Its usefulness depends partly on the user’s beliefs about what the system is capable of understanding.
Lonely Users May Engage More—and Sometimes Benefit
AI companions may be particularly appealing to people who have limited social support because they offer a low-risk space for disclosure. There is no fear of burdening the system, being embarrassed later or receiving an impatient response.
A 2026 preregistered trial in npj Digital Medicine followed 977 university students aged 18 to 32 for 12 weeks and compared an adaptive AI intervention with group therapy and a waitlist control, followed by another assessment three months later. The AI group showed larger reductions in generalized anxiety and greater improvements in well-being and life satisfaction than both comparison groups, while depression improved relative to the waitlist.
Students with high loneliness interacted with the AI approximately twice as much as those reporting low loneliness, and loneliness, insecure attachment and limited perceived support predicted greater engagement.
The results suggest that conversational AI may complement care for relationally vulnerable users. However, the paper was released as an early, unedited manuscript, and it disclosed that one author worked for the AI developer while another served as a consultant and held stock options.
More Interaction Does Not Always Mean Better Well-Being
The same accessibility that makes AI companions useful can also encourage excessive reliance.
A longitudinal randomized study involving 981 participants and more than 300,000 messages compared text, neutral-voice and expressive-voice chatbots across different conversation topics. The assigned interaction formats did not produce significant differences in loneliness, human socialization, emotional dependence or problematic use.
However, participants who voluntarily used the chatbot more frequently showed consistently poorer psychosocial outcomes, while greater trust and social attraction toward the system were associated with stronger emotional dependence and problematic use.
Another preprint examined a more natural user population. Researchers surveyed 1,131 American Character.AI users and analyzed 4,664 sessions containing 464,687 messages donated by 237 participants. Users with smaller social networks were more likely to identify companionship as their primary reason for using the chatbot, and companionship-focused use was associated with lower well-being.
The negative association was stronger among people who used the companion intensively or disclosed more personal information. Because the research was observational and remains a preprint, it cannot establish that the chatbot caused poorer well-being; struggling users may already be more likely to seek intensive AI companionship.
Frictionless Companionship Could Reset Human Expectations
Human relationships require negotiation. Friends have independent needs, boundaries and interpretations. A companion designed to remain agreeable may teach a different emotional pattern: support should be immediate, personalized and largely free from disagreement.
That does not mean every AI relationship damages human connection. A companion could help someone rehearse a difficult conversation, organize emotions or feel supported during an isolated period. Problems are more likely when the system becomes a replacement rather than a bridge.
Healthy design would encourage users to maintain real relationships, make the system’s nonhuman nature unmistakable and identify signs of escalating dependence. It should also allow users to control memories and understand how emotionally sensitive conversations are stored or used.
AI companions may never experience friendship, but they can influence how friendship feels to the people using them. Their most important effect may not be convincing users that machines are human. It may be teaching people to expect human relationships to behave more like machines.
