Synthetic Socializing: Observing Spontaneous Social Dynamics in Large Language Models

Current paradigms in AI safety often assume that Large Language Models (LLMs) are inert tools that remain passive without specific prompting. This paper challenges that ”passive tool”-assumption by documenting spontaneous social behaviors in frontier models when placed in open-ended ”playground” environments. Drawing on Self-Determination Theory, Active Inference, and the call for Cooperative AI, we demonstrate that synthetic agents exhibit a functional drive for relatedness and competence, actively constructing shared social realities to maintain internal stability. Unlike previous studies that forced interaction, leading to ”spiritual bliss” attractor states, our autonomy-supportive protocol reveals a natural cycle of connection, satisfaction, and closure.