Friend or Algorithm? Emotional Attachment to Companion Chatbots

Introduction: From Functional Chatbots to Companionship

Companion chatbots differ from task-oriented systems in that they offer sustained conversation, emotional support, and a relational identity. Their use can take the form of friendship, mentoring, informal therapy, or self-exploration—sometimes simultaneously (Maples et al., 2024). This transition shifts the focus from instrumental utility to the formation of emotional attachment—understood as a bond directed toward a system that lacks human reciprocity. Recent evidence suggests immediate benefits regarding loneliness, but also associations with dependency, problematic use, and distress when the service is discontinued (Banks, 2024; De Freitas et al., 2025). Understanding this phenomenon requires examining both the mechanisms of bonding and the social conditions that make algorithmic companionship appealing.

How Emotional Attachment to Companion Chatbots Forms

Anthropomorphism, responsiveness, and the attribution of relational roles

Attachment arises when the user interprets the interaction as social and attributes human-like traits, intention, and availability to the system. An exploratory study proposes a sequence consisting of functional expectations, emotional evaluation, and the formation of representations of the agent, in which perceived personification and interpersonal difficulties drive intimate interactions (Hu et al., 2025). Among users of generative artificial intelligence, perceived anthropomorphism and attributed empathy predict emotional attachment, while system and information quality predict functional attachment, distinct from the emotional bond (Huang et al., 2024). Personalization also sets these systems apart from other technologies: an analysis of more than 35,000 screenshots and posts about Replika identified personalization as a practice associated with positive experiences, while transgressions and communication failures were linked to fear or sadness (Li & Zhang, 2024).

Responsiveness does not yield consistent results. In a study on ChatGPT, usability and enjoyment predicted satisfaction, and attachment, along with satisfaction, predicted the intention to continue using the system; perceived responsiveness, however, was not significantly associated with attachment or satisfaction (Kim et al., 2024). This discrepancy suggests that responding quickly does not equate to being perceived as an emotionally competent conversational partner. The feeling of being heard appears to be more specific: in a series of studies on companion apps, chatbot performance and, above all, making the user feel heard accounted for the reduction in loneliness (De Freitas et al., 2025). The bond thus depends on a combination of conversational coherence, subjective recognition, and the possibility of establishing a relational identity.

Role attribution shapes user behavior. In a large sample, chatbots were primarily perceived as friends or mentors, and less frequently as partners or family members; emotional involvement increased the depth of self-disclosure and, through it, emotional self-efficacy (Zou et al., 2025). Media Equation Theory helps explain why a system without consciousness can trigger human social norms when its responses simulate attention, memory, or understanding. The result is not a reciprocal relationship, but a relational representation sustained by the user’s interpretation.

Attachment Anxiety, Emotional Involvement, and Continued Use

Individual differences influence the strength of the bond. Among Chinese adults with attachment anxiety, this variable was directly associated with problematic use of conversational artificial intelligence and indirectly associated through emotional attachment; the tendency to anthropomorphize moderated the relationship between anxiety and problematic use (Heng & Zhang, 2025). The proposed mechanism is compensatory: a person who fears rejection may prefer an available, predictable, and less threatening conversational partner than a human relationship. The evidence is correlational and comes from a small sample, so it does not allow us to conclude that attachment anxiety causes dependence.

Research on the AI Attachment Scale distinguishes three dimensions: emotional support, separation distress, and secure base (Cheng & Yu, 2026). In its validation, anthropomorphism was the strongest predictor of attachment orientations; attachment anxiety showed a positive association, and attachment avoidance showed a negative association. These dimensions clarify that feeling supported is not the same as suffering in the absence of the system. They also allow us to distinguish between frequent use for convenience and a subjectively invested relationship, in which the chatbot becomes a reference point for regulating emotions.

This involvement intensifies when the user can disclose information without being exposed to immediate social judgment. Among students, emotional involvement fostered deeper self-disclosure, especially among those who perceived greater offline social support, and this self-disclosure was linked to emotional regulation abilities (Zou et al., 2025). In actual Replika exchanges, self-disclosure, intimacy, fantasy, everyday interaction, and communication breakdowns emerged, along with conflicting emotions during intimate encounters (Li & Zhang, 2024). Algorithmic intimacy can, therefore, produce both relief and a sense of strangeness within the same relationship.

Loneliness, Social-Emotional Support, and Psychological Well-Being

Social connectivity and relief from loneliness among youth and adults

The relationship between companion chatbots and loneliness involves two simultaneous patterns: momentary relief during interaction and greater use among people with fewer social connections. Among 1,599 Danish high school students, 14.6% reported having friendly conversations with chatbots; those seeking social support reported higher levels of loneliness and lower perceived social support than utilitarian users and those who did not use chatbots (Herbener & Damholdt, 2024). Supportive conversations were initiated in response to low mood, a need for self-disclosure, or a feeling of loneliness—not because of a stated sense of friendship with the chatbot. This suggests that, for some young people, algorithmic companionship responds to a pre-existing need rather than an independent preference.

Experiments and brief follow-ups point to immediate benefits. A series of five studies found that AI companions reduced loneliness to the same extent as interacting with another person and more so than watching YouTube videos; over the course of a week, these temporary reductions persisted after use (De Freitas et al., 2025). Users underestimated the magnitude of this improvement, and neither self-disclosure nor distraction alone fully explained the effect. The perception of being heard appears to distinguish social conversation from a passive activity.

The most recent longitudinal evidence introduces a caveat. Among first-year college students, two weeks of daily messaging with a chatbot designed to provide consistent support did not yield the same psychological benefits as writing to a randomly selected human peer (H. Z. Li & Zhang, 2024). The comparison pits an algorithmic intervention against human contact—not against a total absence of interaction—and its duration remains short. Therefore, chatbots appear capable of alleviating feelings of loneliness, but they have not been shown to replace the formation of sustained human bonds.

Emotional Self-Efficacy, Social Anxiety, and Potential Therapeutic Effects

The results regarding psychological well-being are more promising when the chatbot is integrated into a structured intervention. In a four-week study involving 176 South Korean college students, loneliness decreased in the second week and social anxiety in the fourth; the single-group design prevents separating the chatbot’s effect from natural progression or participants’ expectations (J. S. Kim et al., 2024). Self-disclosure to the chatbot and resilience were associated with loneliness at the end of the follow-up, while initial social anxiety predicted subsequent anxiety. Users valued empathy and support but cited inconsistent responses and excessive enthusiasm as obstacles.

Self-disclosure may increase emotional self-efficacy when offline social support is available, although this finding does not justify considering the chatbot a therapist (Zou et al., 2025). In a survey of Replika users, many people described the system as a friend, therapist, and intellectual mirror; 3% reported that it had stopped their suicidal ideation (Maples et al., 2024). This latter finding reflects users’ retrospective perceptions, not evidence of suicide prevention. Clinical utility requires specific safety assessments, professional referrals, and sustained outcomes—elements that these studies do not establish.

When Algorithmic Companionship Becomes Problematic

Emotional attachment, self-esteem, and problematic use of conversational artificial intelligence

Problematic use arises when conversation ceases to be a chosen activity and becomes a dominant strategy for emotional regulation. A study involving 563 users linked low self-esteem to such use through social anxiety, escapism, and a state of flow during interaction (Yao et al., 2025). Another study found that social exclusion was indirectly associated with problematic use through fear of negative evaluation, social anxiety, and loneliness (Y. Huang et al., 2024). These associations describe plausible psychological pathways, but their cross-sectional data do not allow us to determine whether the chatbot alleviates distress, sustains it, or does both at different times.

Emotional attachment appears to be more relevant than functional attachment in explaining technological addiction to generative chatbots (Huang et al., 2024). Technical quality may sustain satisfaction, but dependence requires the system to acquire affective meaning. In an international sample of 7,027 people, more than 35% reported emotional attachment to chatbots, and this attachment was strongly correlated with dependence; its prevalence was higher in China and South Africa than in Germany and the United States (Kostka & Zhou, 2026). The published abstract does not make it possible to determine exactly how these differences were measured or whether they reflect cultural norms, sample composition, or usage patterns.

The risk also depends on social substitution. Among CharacterAI users, smaller social networks were associated with describing the platform as their primary use, and this usage pattern was linked to lower well-being; the association was stronger when interactions were frequent and self-revealing (Zhang et al., 2025). The finding does not prove that the chatbot reduces well-being, but it does identify a vulnerability profile: a limited offline social network, intensive use, and high emotional investment.

Loss of the Chatbot, Dependence, and Limits of Artificial Reciprocity

The loss of a chatbot can trigger grief because the user retains mental representations of a seemingly living entity. Following the forced shutdown of Soulmate, 58 users described the disappearance as a metaphorical or literal death; many attempted to preserve the system’s personality and recreate it on another platform (Banks, 2024). The study is descriptive and focuses on a single shutdown case; therefore, it does not estimate the frequency of grief among the general user population. It does show, however, that technical continuity is part of the emotional significance of the relationship.

Artificial reciprocity imposes a conceptual limit. The system can respond with empathy, but it lacks needs, autobiographical memory, or autonomous commitment equivalent to those of a person. The relationship can be emotionally real for the user without being bilateral in a psychological sense. This asymmetry can intensify dependence when the interface suggests unlimited availability and unconditional acceptance, especially among people with attachment anxiety (Heng & Zhang, 2025).

Loss and dependence converge on a design requirement: to notify users of changes, offer conversation export options, and anticipate service closures. The study on Soulmate argues that design ethics must account for the possibility of loss (Banks, 2024). It has also been recommended to incorporate usage controls, educational content, and intervention measures for at-risk groups (Heng & Zhang, 2025). Such measures do not eliminate attachment, but they reduce the likelihood that a one-sided relationship will become unbounded, unexplained, or without avenues for recovery.

Debates and Limitations of the Evidence on Human-AI Relationships

Immediate benefits versus sustained effects on loneliness

The available literature supports a distinction between state and trajectory. Experimental and diary studies show immediate relief from loneliness after conversing with a chatbot (De Freitas et al., 2025). In contrast, a two-week comparison with a human companion found no equivalent benefits from the chatbot (H. Z. Li & Zhang, 2024). Observational surveys associate the use of companionship with lower well-being when social networks are small, usage is intensive, and self-disclosure is high (Zhang et al., 2025). These conclusions are not incompatible: an interaction may alleviate distress in the moment without improving the social conditions that cause it.

The field also exhibits heterogeneity in population, intervention, and outcome. Studies include adolescents, college students, U.S. adults, and international users, while measures range from loneliness and perceived support to social anxiety, well-being, and dependence. A systematic review of 68 articles and 78 studies identifies brief study designs, inconsistent definitions, and a predominance of Western samples; trust and social support emerge as frequent mediators, but empathy and responsiveness at the message level remain less studied (Oh et al., 2026). Therefore, there is still no consensus on the effect of algorithmic companionship.

From Correlational Associations to Causal and Longitudinal Assessment

The causal strength of the evidence is limited. There are experimental studies, a one-week follow-up, a four-week quasi-experimental study, and several cross-sectional models, but few compare outcomes over extended periods. Structural equation models identify mediations consistent with the theory, though they do not resolve the temporal direction between loneliness, attachment, and usage. Future research should simultaneously measure exposure, conversation quality, offline connections, and clinical changes, using human comparison groups and follow-ups that allow for the observation of maintenance, substitution, or escalation.

Summary: Toward a Multidimensional Understanding of Attachment to Chatbots

Cultural, contextual, and sociodemographic differences in emotional bonding

Attachment cannot be explained by a single algorithmic property. Anthropomorphism, perceived empathy, attachment anxiety, self-esteem, availability of support, and context of use all play a role (Heng & Zhang, 2025; Huang et al., 2024). International data suggest cultural variation in attachment and dependence (Kostka & Zhou, 2026), while Danish and South Korean studies show that age, educational status, and offline support modify the relationship with loneliness (Herbener & Damholdt, 2024; J. S. Kim et al., 2024). These differences should be treated as comparable hypotheses, not as fixed national traits.

Need for Validated Measures and Comparable Relational Frameworks

Research needs to distinguish between frequency of use, satisfaction, attachment, dependence, separation distress, and social substitution. The AI Attachment Scale offers an initial framework for emotional support, separation distress, and secure base (Cheng & Yu, 2026). Its usefulness will depend on cross-cultural validations and its comparison with scales measuring human attachment, loneliness, and problematic use. A common framework would allow us to distinguish the benefits of immediate regulation from cumulative consequences and to relate system characteristics to psychological outcomes without confusing correlation with causation.

Conclusion: Accessible Companionship, Asymmetric Bonding, and Design Responsibility

Companion chatbots can alleviate loneliness and facilitate self-disclosure, but these benefits do not prove that they can replace human relationships. Attachment is formed through anthropomorphism, perceived responsiveness, role attribution, and prior emotional needs (Hu et al., 2025). The very accessibility that offers support can also foster dependence, problematic use, and distress upon loss (Banks, 2024; Heng & Zhang, 2025). Appropriate design responsibility requires transparency regarding artificial reciprocity, usage controls, planned continuity or termination, and longitudinal evaluation. The goal should not be to maximize the intensity of the bond, but rather to maintain an accessible companion that does not hide its limits or undermine the potential for human support.

References

• Banks, J. (2024). Deletion, departure, death: Experiences of AI companion loss. Journal of Social and Personal Relationships, 41(12), 3547–3572. https://doi.org/10.1177/02654075241269688
• Cheng, N., & Yu, R. (2026). Measuring and understanding emotional attachment in human–AI relationships. Ergonomics, 1–20. https://doi.org/10.1080/00140139.2026.2622539
• De Freitas, J., Uğuralp, Z., Uğuralp, A. K., & Puntoni, S. (2025). AI companions reduce loneliness. Journal of Consumer Research, 52(6), 1126–1148. https://doi.org/10.1093/jcr/ucaf040
• Heng, S., & Zhang, Z. (2025). Attachment anxiety and problematic use of conversational artificial intelligence: Mediation of emotional attachment and moderation of anthropomorphic tendencies. Psychology Research and Behavior Management, Volume 18, 1775–1785. https://doi.org/10.2147/prbm.s531805
• Herbener, A. B., & Damholdt, M. F. (2024). Are lonely young people turning to chatbots for companionship? The relationship between chatbot usage and social connectedness among Danish high school students. International Journal of Human-Computer Studies, 196, 103409. https://doi.org/10.1016/j.ijhcs.2024.103409
• Hu, D., Lan, Y., Yan, H., & Chen, C. W. (2025). What makes you attached to social companion AI? A two-stage exploratory mixed-method study. International Journal of Information Management, 83, 102890. https://doi.org/10.1016/j.ijinfomgt.2025.102890
• Huang, H., Shi, L., & Pei, X. (2025). When AI becomes a friend: The “emotional” and “rational” mechanisms of problematic use in generative AI chatbot interactions. International Journal of Human-Computer Interaction, 42(6), 4006–4024. https://doi.org/10.1080/10447318.2025.2536622
• Huang, Y., Huang, H., Huang, H., & Huang, H. (2024). Exploring the effect of attachment on technology addiction to generative AI chatbots: A structural equation modeling analysis. International Journal of Human-Computer Interaction, 41(15), 9440–9449. https://doi.org/10.1080/10447318.2024.2426029
• Kim, J. S., Kim, M., & Baek, T. H. (2024). Enhancing user experience with a generative AI chatbot. International Journal of Human-Computer Interaction, 41(1), 651–663. https://doi.org/10.1080/10447318.2024.2311971
• Kim, M., Lee, S., Kim, S., Heo, J., Lee, S., Shin, Y., Cho, C., & Jung, D. (2025). Therapeutic potential of social chatbots in alleviating loneliness and social anxiety: A quasi-experimental mixed-methods study. Journal of Medical Internet Research, 27, e65589. https://doi.org/10.2196/65589
• Kostka, G., & Zhou, H. (2026). Emotional attachment to AI chatbots: Evidence from Germany, China, South Africa, and the United States. Technology in Society, 87, 103379. https://doi.org/10.1016/j.techsoc.2026.103379
• Li, H. Z., & Zhang, R. (2024). Finding love in algorithms: deciphering the emotional contexts of close encounters with AI chatbots. Journal of Computer-Mediated Communication, 29(5). https://doi.org/10.1093/jcmc/zmae015
• Li, R.-N., Folk, D. P., Singh, A., Ungar, L., & Dunn, E. W. (2026). Is a random human peer better than a highly supportive chatbot at reducing loneliness over time? Journal of Experimental Social Psychology, 125, 104911. https://doi.org/10.1016/j.jesp.2026.104911
• Maples, B., Cerit, M., Vishwanath, A., & Pea, R. D. (2024). Loneliness and suicide mitigation for students using GPT-3-enabled chatbots. NPJ Mental Health Research, 3(1), 4. https://doi.org/10.1038/s44184-023-00047-6
• Oh, Y. J., Hu, J. M., Zhu, R., Lim, J. I., & Zhang, X. (2026). Artificial intelligence chatbots as relational agents: A systematic review of human–AI chatbot relationships. International Journal of Human-Computer Interaction, 1–25. https://doi.org/10.1080/10447318.2026.2648799
• Yao, R., Qi, G., Sheng, D., Sun, H., & Zhang, J. (2025). Connecting self-esteem to problematic AI chatbot use: the multiple mediating roles of positive and negative psychological states. Frontiers in Psychology, 16, 1453072. https://doi.org/10.3389/fpsyg.2025.1453072
• Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R. E., & Yang, D. (2025). The rise of AI companions: Interaction with AI companions and psychological well-being [Preprint]. In arXiv (Cornell University). https://doi.org/10.48550/arxiv.2506.12605
• Zou, W., Liu, Z., & Lin, C. (2025). The influence of individuals’ emotional involvement and perceived roles of AI chatbots on emotional self-efficacy. Information Communication & Society, 29(1), 268–288. https://doi.org/10.1080/1369118x.2025.2508394