Abstract: 【Purpose/significance】Amid the increasing frequency of public emergencies, Artificial Intelligence Generated Content (AIGC) has become deeply integrated into social media networks, influencing users' information behaviors and accelerating the shift from User Generated Content (UGC) to a "human-AI collaboration" paradigm. This study focuses on the interactive behaviors between human users and AI agents on social media, aiming to uncover the characteristics and mechanisms of human-AI interactions and to provide insights for emergency governance in the intelligent era.【Method/process】Using Python and Botometer, the study collected and identified comments from human users on tweets posted by AI agents on the X platform (formerly Twitter) during public emergency outbreaks. Adopting the three-level coding approach of classic grounded theory, the research systematically analyzed these comments and constructed a model of social media users' human-AI interaction behaviors in public emergencies.【Result/conclusion】The find⁃ ings reveal that: (1) in the cognitive dimension, users display risk awareness, information judgment, and social cognition orientation; (2) in the emotional dimension, users experience "emotional triggers" that lead to "emotional venting", "empathy" and "expressions of gratitude and encouragement"; (3) in the behavioral dimension, users exhibit "information sharing" "information seeking" "factchecking," and "informational conformity" tendencies. Overall, AI agents significantly influence human users’cognitive processing, with human-AI interactions demonstrating high information sensitivity and emotional contagion. These findings reveal a dynamic cou⁃ pling of cognition, emotion, and behavior that shapes complex, evolving interaction patterns, offering practical implications for platform governance and emergency management.【Innovation/limitation】This study introduces the "human-AI interaction" concept into emer⁃ gency management, establishing a three-dimensional cognition–emotion–behavior model that reveals AI's impact on public informa⁃ tion behavior. However, its generalizability is constrained by platform-specific data, linguistic context, and qualitative methodological limitations, suggesting the need for future multi-platform and longitudinal validation.