This study investigated the potential of Large Language Models (LLMs), like GPT-4, in generating roles that simulate human social behavior. By combining both quantitative and qualitative analyses, the study assessed the effectiveness of these models in role creation and simulating complex human interactions. Specifically, the model generated 50 fictional roles, each with distinct personality traits and backgrounds, which then interacted with real users on a simulated social platform over two weeks. We obtain three main observations: 1) These roles exhibited a broad range of personality traits and diverse backgrounds, earning high ratings for emotional authenticity; 2) The interactions demonstrate the model's advanced capabilities in simulating complex social networks and dynamics, mirroring the diversity and complexity of real social environments. 3) The model has limited ability to handle extreme emotional states and complex social scenarios. The findings have significant implications for digital entertainment, education, and social media, providing valuable insights into the application of LLMs in creating virtual roles.