Abstract
Emotions play a central role in shaping individuals' cognitive responses and selfexpression, while AI chatbots are emerging as novel mediators for psychological support and self-exploration. However, existing research has often simplified human-machine interactions into linear cognitive processes, overlooking the underlying nonlinear and necessary psychological mechanisms. This study constructs a multilayered model integrating emotional arousal, affective engagement, and psychological acceptance to reveal how different interaction modes (text, voice, and multimodal) influence users' self-disclosure behaviors. Based on 352 valid survey responses, the study employed Structural Equation Modeling (SEM) to validate the core path relationships and further integrated Artificial Neural Network (ANN) and Necessary Condition Analysis (NCA) to uncover nonlinear effects and necessity thresholds among latent variables. Results indicated that voice and multimodal interactions significantly enhanced users' emotional arousal and psychological acceptance, both of which served as key mechanisms facilitating self-disclosure. Moreover, the ANN analysis revealed the non-compensatory nature of interaction mode effects, while the NCA results further identified emotional arousal and acceptance as indispensable conditions for high levels of self-disclosure. These findings suggest that AI chatbots should not be viewed merely as tools for information exchange but rather as co-constructs of emotion and trust, whose interaction design should focus on the dynamic balance between affective engagement and psychological safety. This study provides a novel theoretical perspective for understanding human-AI emotional resonance and psychological expression, as well as design implications for developing AI intervention systems with enhanced psychological sensitivity.