TLDR: A longitudinal study reveals how AI companionship develops, moving beyond human-likeness to highlight the critical roles of an AI’s perceived agency, parasocial interaction, and user engagement. Through a survey of 303 users and a four-week study with a generic chatbot, researchers found that users’ perceptions of a new AI companion quickly converge with their existing AI relationships. The study validates a comprehensive psychological pathway for AI companionship and proposes a new empirical method for research, emphasizing the need for responsible design that considers rapid relationship formation and potential psychological impacts.
AI companions are rapidly gaining popularity, but their widespread use also presents significant questions regarding mental health, personal well-being, and social relationships. While previous research has identified various individual factors influencing human-companion interactions, a comprehensive understanding of how these factors interact and evolve over time has been largely missing. A recent study, titled How AI Companionship Develops: Evidence from a Longitudinal Study, by Angel Hsing-Chi Hwang, Fiona Li, Jacy Reese Anthis, and Hayoun Noh, addresses this gap by offering an integrated view of how human-AI companionship forms and progresses.
The researchers conducted two main studies. Study 1 involved a large-scale survey of 303 existing AI companion users (e.g., Replika, Character.ai). The goal was to map the psychological journey from users’ initial mental models of the AI agent to their parasocial experiences, social interaction patterns, and the overall psychological impact of these companions. This initial study revealed that multiple interconnected variables—specifically, the AI’s perceived agency (its capacity to act), the user’s parasocial interaction (a one-sided, perceived relationship), and the user’s engagement—are crucial in shaping AI companionship.
Building on these findings, Study 2 was a longitudinal investigation with a subset of 110 participants from Study 1. These participants interacted with a new, generic chatbot, dubbed “Study Bot,” once a week for four weeks. This part of the research aimed to see if interactions in a controlled empirical setting could simulate real-world experiences with personal AI companions. A remarkable finding was that participants’ perceptions of the generic Study Bot significantly converged with their perceptions of their own AI companions by the third week of interaction. This suggests that human-AI companionship can develop rapidly, even with a basic, design-agnostic chatbot.
The study’s contributions are threefold. Firstly, it distills and empirically validates a general pathway of theoretical constructs that explain human-AI companionship dynamics, many of which were previously studied in isolation. This pathway moves from a user’s mental model of the agent, through parasocial experience and social penetration, to the psychological impact of the AI companion.
Secondly, the research highlights a set of interconnected factors beyond just human-likeness that shape AI companionship. These include the AI’s technical capabilities and responsiveness, which have often been underemphasized in prior research. The findings suggest that companionship is not solely about how human-like an agent appears, but also how effectively it can respond and engage with the user.
Thirdly, the study introduces a practical methodology for future human-computer interaction (HCI) research. By demonstrating that users’ responses to their own AI companions can be replicated through weekly interactions in controlled settings, the researchers offer a way to gain insights without relying on often inaccessible data from commercial service providers. This is particularly valuable given the challenges of obtaining real-world, longitudinal data on AI companion use.
Theoretical and Design Insights
The research offers several key implications for understanding and designing AI companions. Theoretically, it emphasizes the central role of “agency” (the AI’s capacity to act and respond) and “parasocial experience” (the feeling of a one-sided relationship) in forming companionship. It challenges the notion that anthropomorphism is the sole or primary driver, suggesting that an AI’s ability to present itself as an autonomous and responsive entity is equally, if not more, important.
Unique disclosure patterns also emerged. While self-disclosure is typically seen as positive in human relationships, with AI companions, it can heighten dependence and over-reliance. Interestingly, the sentiment of disclosure (positive vs. negative) did not significantly mediate relationship deepening, suggesting that sustained engagement, regardless of emotional valence, might be a key factor.
From a design perspective, the study suggests rethinking the necessity of overly anthropomorphic designs, as even generic chatbots can elicit strong psychological impacts. Designers are encouraged to calibrate user perceptions, perhaps by transparently reminding users of the AI’s non-human nature, especially after intense interactions. The study also highlights “time” as a critical design variable, noting that companionship forms quickly. This calls for slower onboarding processes and adaptive systems that can modulate responsiveness over time to prevent over-reliance. Finally, the importance of designing robust “disengagement pathways” is stressed, providing users with mechanisms to take breaks, connect with human networks, or access professional resources if problematic reliance patterns emerge.
Also Read:
- Mapping the Landscape of Human–AI Trust in Generative AI
- Navigating Mental Health AI: A Framework for Safer Disclosure and Enhanced User Understanding
Limitations and Future Directions
The authors acknowledge several limitations, including the exploratory nature of the study, which limits strong causal inferences. The online setting also constrains ecological validity, meaning findings might not fully capture real-world complexities. Self-selection bias, where more engaged participants were likely to complete the longitudinal study, could exaggerate observed effects. The generalizability is also limited to users already familiar with AI companions and specific demographic profiles. Future work should address these by adopting experimental designs, conducting field-based studies, minimizing attrition, and examining diverse populations and cultural contexts.
In conclusion, this research provides a crucial, integrated framework for understanding the dynamic development of AI companionship. It moves beyond fragmented views, offering theoretical insights and methodological approaches that can guide the responsible design of AI companions, ensuring they complement human well-being rather than displacing essential social and healthcare infrastructures.
