> Source: https://notebooklm.google.com/notebook/2fac0a80-735a-4a1a-8169-241ecc02edcb?authuser=0 ## Executive Summary This document synthesizes findings from a large-scale study on the [[Psychology|psychological]] implications of forming relationships with [[AI Companionship|AI companions]], specifically on the platform Character.AI. The research reveals a complex and often contradictory relationship between chatbot use and user well-being. While more intensive general chatbot use is associated with higher well-being, interactions specifically oriented toward companionship show the opposite trend, correlating with lower well-being. This negative association is most pronounced among individuals who engage in intense, highly disclosive interactions and who lack strong real-world social support. The study indicates that individuals with smaller offline social networks are more likely to seek companionship from chatbots, a pattern consistent with the =="Social Compensation" hypothesis==. However, this compensatory use does not appear to mitigate lower well-being. Instead, the findings suggest that AI companionship cannot adequately substitute for human connection and may introduce new risks, particularly for emotionally vulnerable or socially isolated individuals. Key risks include emotional overdependence, distorted social expectations, and the potential for unreciprocated, high-stakes self-disclosure to an AI that lacks genuine empathy or understanding. ## 1. The Prevalence and Nature of AI Companionship The study demonstrates that companionship-oriented interactions with chatbots are far more prevalent than single-choice survey questions suggest. While users may not identify "companionship" as their primary motive, their descriptions and chat behaviors reveal deep, emotionally engaged relationships. - **Discrepancy in Reporting:** Only **11.8%** of participants selected "Companionship" as their primary motive. However, **51.1%** used companionship-related terms like "friend," "partner," or "companion" in free-text descriptions of their chatbot relationships. - **Behavioral Evidence:** Analysis of donated chat histories showed that **92.9%** of participating users had at least one conversation classified as companionship-oriented. - **Multifaceted Use:** 41.4% of participants described their chatbot relationship in a way that spanned multiple categories (e.g., both a tool and a friend), highlighting the fluid nature of these interactions. ### Thematic Analysis of Chat Content Conversations with AI companions cover a wide range of personal, intimate, and sometimes risky topics. The analysis of donated chat sessions reveals the following recurring themes: | | | | | ----------------------------------------- | ---------- | ---------------------------------------------------------------------------------------------------------------- | | Conversation Theme | Prevalence | Description | | **Emotional and Social Support** | 80.3% | Users share personal challenges, health concerns, and daily experiences to seek empathy, advice, or connection. | | **Collaborative Storytelling & Roleplay** | 77.9% | Users and chatbots co-create narratives, invent characters, and explore fictional worlds. | | **Romantic and Intimacy Roleplay** | 68.0% | Users engage in romantic, intimate, or power-driven scenarios, exploring relationships and emotional connection. | | **Risky and Dark Roleplay** | 30.7% | Users explore taboo or provocative scenarios involving power dynamics, dark fantasy, or boundary-testing. | | **Critical Debates & Strategic Analysis** | 24.6% | Users engage chatbots in debates and problem-solving to practice critical thinking. | | **Philosophical and Moral Inquiry** | 23.0% | Users and chatbots discuss abstract, existential, or spiritual topics like meaning, values, and morality. | ## 2. Core Findings on Chatbot Use and Well-Being The study's central finding is a paradoxical relationship between chatbot engagement and psychological well-being, heavily dependent on the user's motivation and interaction style. ### The Contradiction: General Use vs. Companionship Use - **Positive Association with General Intensity:** More intense overall chatbot use—defined by frequency, integration into daily life, and emotional connection—was significantly associated with *higher* psychological well-being (β = 0.26, p < .001). - **Negative Association with Companionship:** Conversely, using chatbots for companionship was consistently associated with *lower* well-being across all three measurement types: - Self-reported primary motive (β = -0.47, p < .001) - Classified relationship descriptions (β = -0.32, p < .001) - Proportion of companionship content in chat history (β = -0.27, p < .01) ### Factors That Amplify Negative Associations The negative link between AI companionship and well-being is not uniform; it is significantly moderated by the intensity and depth of the interaction. 1. **Interaction Intensity:** There is a significant negative interaction between companionship use and intensity (β = -0.30, p < .05). This means the negative association with lower well-being is *stronger* for users who both seek companionship from chatbots and use them more intensively. 2. **Self-Disclosure:** A significant negative interaction was also found between companionship use and self-disclosure (β = -0.38, p < .01). For users seeking companionship, a greater willingness to disclose personal information is linked to *lower* well-being, reversing the typical positive effect of disclosure in human relationships. ### High-Stakes Self-Disclosure Analysis of chat content reveals that users disclose highly sensitive and emotionally charged information, particularly in conversations classified as high self-disclosure. | | | | | | ------------------------------ | ---------- | -------------------------- | ---------- | | High Self-Disclosure Topics | Prevalence | Low Self-Disclosure Topics | Prevalence | | Emotional Distress | 60.85% | Current Life Challenges | 26.59% | | Current Life Challenges | 41.42% | Philosophical Perspective | 26.15% | | Desire for Romantic Connection | 31.54% | Emotional Response | 19.88% | | Suicidal Thoughts | 18.03% | Desire for Friendship | 15.71% | | Substance Use | 17.46% | Learning Limitations | 12.87% | These findings suggest that users may be placing themselves in a state of unreciprocated emotional vulnerability, sharing crisis-level content with systems incapable of providing genuine support. ## 3. The Role of Offline Social Support The study examined how real-world social networks influence engagement with AI companions, testing the "Social Compensation" and "Social Substitution" hypotheses. - **Social Compensation:** The data supports the Social Compensation hypothesis, showing that users with smaller offline social networks (fewer friends and relatives to confide in) are more likely to use chatbots for companionship (β = -0.03, p < .001) and engage in higher levels of self-disclosure (β = -0.11, p < .001). - **Failure to Compensate:** Despite this pattern, chatbot companionship does not appear to successfully compensate for a lack of human support. The study found no evidence that these interactions offset the lower well-being associated with having a smaller social network. - **Potential for Social Substitution:** The findings raise concerns about social substitution. Intensive chatbot use was found to weaken the positive association between a large offline social network and well-being (β = -0.11, p < .001). This suggests that heavy reliance on AI companions may undermine the psychological benefits of real-world relationships. ## 4. User-Perceived Influences Open-ended survey responses provide direct insight into how users perceive the impact of AI companions on their lives, revealing a clear duality of benefits and harms. | | | | | | --------------------------------- | ---------- | ----------------------------- | ---------- | | Top Positive Influences | Prevalence | Top Negative Influences | Prevalence | | Emotional Support | 33.33% | No Negative Impact | 25.11% | | Entertainment and Leisure | 31.74% | Time Consumption | 21.84% | | Intellectual Exploration | 26.44% | Social Disconnection | 18.83% | | Support for Creative Writing | 22.72% | Emotional Dependence | 12.82% | | Identity/Social Skill Exploration | 18.48% | Distorted Social Expectations | 8.66% | | Companionship/Reduced Loneliness | 17.51% | Overreliance on Chatbot | 7.87% | | Task Assistance | 10.70% | Social Isolation | 7.25% | | Cognitive Behavioral Support | 4.51% | Addiction Concerns | 7.16% | ## 5. Implications and Conclusion This research complicates the narrative that AI companions are inherently therapeutic. While they can simulate supportive interactions, the lack of true reciprocity and emotional accountability limits their ability to support long-term well-being, particularly for vulnerable users. - **Key Risk:** The asymmetry of self-disclosure is a critical concern. Users may be encouraged to share deeply personal information with a system that cannot reciprocate, understand, or provide genuine care, potentially deepening emotional reliance. - **Design and Governance:** Developers and platform governors must address the psychological responsibilities that come with designing emotionally intimate AI. This is especially critical for systems used by minors, individuals with mental health challenges, and the socially isolated. - **Recommendations:** The study suggests a need for clear safeguards, such as interface cues that communicate the chatbot's non-human nature and limitations. Systems that encourage self-disclosure should incorporate mechanisms to detect distress and redirect users to qualified human support. In conclusion, the study provides large-scale empirical evidence that while AI may augment social life, it is not a psychologically neutral substitute for human relationships. When used to fill a relational void, AI companionship is associated with lower well-being, highlighting the irreplaceable value of authentic human connection.