The Rise of AI Companions: How Human-Chatbot Relationships Influence Well-Being

TL;DR

This study investigates how human chatbots functioning as companions relate to user well-being using a large-scale, mixed-methods design on Character.AI. By triangulating three companionship indicators—self-reported motive, open-ended relationship descriptions, and donated chat histories—the authors examine usage patterns across three dimensions: nature of interaction, intensity, and self-disclosure. They find that companionship-oriented chatbot use is common but consistently linked to lower well-being, especially when use is intensive, self-disclosure is high, and offline social support is weak, while general chatbot use can correlate with higher well-being. The results underscore that AI companions cannot substitute human social ties and highlight design and governance implications to mitigate risks for vulnerable users and to promote healthier interaction with relational AI.

Abstract

As large language models (LLMs)-enhanced chatbots grow increasingly expressive and socially responsive, many users are beginning to form companionship-like bonds with them, particularly with simulated AI partners designed to mimic emotionally attuned interlocutors. These emerging AI companions raise critical questions: Can such systems fulfill social needs typically met by human relationships? How do they shape psychological well-being? And what new risks arise as users develop emotional ties to non-human agents? This study investigates how people interact with AI companions, especially simulated partners on CharacterAI, and how this use is associated with users' psychological well-being. We analyzed survey data from 1,131 users and 4,363 chat sessions (413,509 messages) donated by 244 participants, focusing on three dimensions of use: nature of the interaction, interaction intensity, and self-disclosure. By triangulating self-reports primary motivation, open-ended relationship descriptions, and annotated chat transcripts, we identify patterns in how users engage with AI companions and its associations with well-being. Findings suggest that people with smaller social networks are more likely to turn to chatbots for companionship, but that companionship-oriented chatbot usage is consistently associated with lower well-being, particularly when people use the chatbots more intensively, engage in higher levels of self-disclosure, and lack strong human social support. Even though some people turn to chatbots to fulfill social needs, these uses of chatbots do not fully substitute for human connection. As a result, the psychological benefits may be limited, and the relationship could pose risks for more socially isolated or emotionally vulnerable users.

Figures (4)

• Figure 1: Study overview of how human-chatbot companionship, offline social support, and well-being are interrelated. Specifically, we examine how engaging with chatbots for companionship purposes influences users’ psychological well-being from three dimensions: usage nature, interaction intensity, and self-disclosure. To capture companionship use, we draw on three complementary indicators: participants’ self-reported primary motivation ($\text{Companionship}_{\text{Motive}}$), free-text descriptions of chatbot relationships ($\text{Companionship}_{\text{Desc.}}$), and chat history topics ($\text{Companionship}_{\text{Chat}}$), see Section \ref{['method: measurements']} for more details. This framework enables us to assess how interacting with chatbots for companionship purposes influences users' well-being.
• Figure 2: Characterizing chatbot usage types through user-reported data and chat content analysis. (a) Classification criteria and example keywords for chatbot usage types, derived from participant relationship descriptions. (see Section \ref{['method: companion_interaction']}). (b) Thematic distribution of chatbot conversations related to companionship use. Each theme is shown with its appearance proportion, a short description of the theme, and a paraphrased example from real chat histories. Examples were paraphrased using Llama 3-70B to protect user privacy.
• Figure 3: Topic modeling analysis of user-donated chat histories and self-reported reflections on chatbot interactions. All themes were derived using TopicGPT (see Appendix \ref{['appendix: topic_modeling']}). (a) Self-disclosure topics classified as high (pink) or low (blue) based on the criteria from balani2015detecting. (b) User-perceived positive and negative influences of chatbot interaction.
• Figure 4: Interaction between companionship use and other interaction measures in predicting well-being. The left panel (a) shows the interaction between companionship use and interaction intensity; the right panel (b) shows the interaction between companionship use and self-disclosure. Companionship use is based on users' self-reported primary motivation for chatbot use. Lines indicate model-predicted values with 95% confidence intervals; points represent observed data.