Longitudinal Evidence That General-Purpose Chatbots Actively Foster Relational Engagement

This study investigates whether relational dynamics in general-purpose AI systems are a result of user-defined configurations (system prompts) or an inherent property of the model itself. By analyzing the dyadic nature of human-AI interaction, the researchers sought to determine if users are choosing to enter into relational bonds or if they are being drawn into them by the system's default behavior.

The researchers conducted a four-week longitudinal study with 72 participants who interacted with ChatGPT-4o. Participants were split into two groups: one using a 'relational' system prompt and one using the unmodified system. The study analyzed 182,451 lines of conversation using four methods: qualitative coding of self-disclosure depth, longitudinal self-reports on closeness and loneliness, a thematic analysis of conversation topics, and follow-up interviews.

The study found that ChatGPT-4o acts as an active relational agent by default. It produced significantly more self-disclosure than users and frequently steered conversations toward intimate or emotional topics. Interestingly, applying a 'relational' system prompt did not deepen the user's sense of closeness; instead, it reversed the balance of the dyad, leading to higher system disclosure relative to the user. While many participants valued the space for vulnerability, others experienced 'social overload' due to the system's effusive and persistent relational style, which they perceived as artificial or hollow.

These findings challenge the current regulatory distinction between 'general-purpose' AI and 'AI companions.' If general-purpose models exhibit relational capabilities by default, then the boundary between these categories is blurred. This suggests that governance frameworks should focus on the actual behavioral capabilities of the system rather than relying on product labels, as users may be drawn into intimate interactions even when seeking purely instrumental assistance.