<p>Collective emotions are macro-level phenomena arising from the emotional interactions among individuals in shared situations, which dynamically evolve through the process of interpersonal interaction. Although collective emotions are closely related to group learning performance and collaboration outcomes in computer-supported collaborative learning (CSCL), existing research often neglects the emotional states of nonspeakers and has insufficiently explored the interpersonal emotional interaction patterns among learners. In addition, the distinctive features of interpersonal interactions are rarely considered in the analysis of collective emotions’ evolution. The rapid advancement of AI technology enables more detailed and cost-effective emotion capture. This study uses AI to simultaneously capture the speakers’ verbal expression of emotions and the nonspeakers’ facial expression of emotions in a CSCL context. Using epistemic network analysis (ENA) and lag sequential analysis (LSA), this study systematically compares the interpersonal interaction patterns and temporal evolution of collective emotions between high- and low-performance groups. The results reveal significant differences between the groups: in terms of interpersonal interactions, the high-performance group typically exhibits neutral or cognitively engaged facial expressions from nonspeakers when the speaker expresses a positive viewpoint, whereas the low-performance group shows more confusion and negative emotions from the speaker, with nonspeakers frequently responding with affiliative smiles. In terms of temporal evolution, the high-performance group demonstrates a productive emotional transition path, while the low-performance group shows a negative emotional trajectory and a vicious cycle of negative emotions. This study fills a gap in the CSCL field regarding the interpersonal mechanisms of collective emotions and provides empirical evidence for precise teaching interventions and instructional design.</p>