Light Trumps Shadow? How Generative AI Agent’s Language Arousal Influences Users’ Interactive Willingness: Evidence From Multimodal Analysis

This study investigates how language arousal in Generative AI systems influences users' interaction willingness, examining the roles of social identity and visual atmosphere. Drawing on the Limited Capacity Model of Motivated Mediated Message Processing (LC4MP) and Social Identity Theory, we constructed a theoretical model integrating language arousal, social identity, visual atmosphere, and interaction willingness, and analyzed 8,809 interactions from Character.AI using multimodal methods combining linguistic analysis and visual processing. Our findings reveal that high-arousal language significantly increases interaction willingness, with social identity mediating this relationship. Most notably, we discovered a “psychological defense-curiosity paradox”: shadow visual atmospheres, despite triggering initial defensive reactions, enhance engagement more effectively than light atmospheres, challenging conventional “brighter is better” design assumptions. This research advances theory by repositioning language arousal as a direct causal variable in AI interaction, extending cognitive processing models to human-AI contexts, and demonstrating how visual elements strategically modulate psychological responses. These insights provide valuable direction for developing emotionally intelligent AI systems that effectively balance linguistic stimulation and visual atmosphere to create more engaging human-AI experiences.