The rapid integration of generative Artificial Intelligence (GAI) into human–AI collaborative learning contexts has attracted extensive attention from educators and researchers. However, relatively little is known about the underlying cognitive and neural evidence regarding the differential effects of two instructional paradigms—Teacher-Student Interaction (TSI) and GAI-Student Interaction (GSI)—on students’ learning engagement and academic performance. This study adopted a within-subjects experimental design to compare students’ performance in science learning under TSI and GSI conditions, and used functional near-infrared spectroscopy (fNIRS) hyperscanning to investigate the associated cognitive neural mechanisms. The results indicated that GSI significantly enhanced students’ cognitive engagement, whereas TSI more effectively fostered behavioural, emotional, and social engagement. These effects were impacted by students’ metacognitive levels across both instructional paradigms. Neural evidence further demonstrated that TSI was characterised by reciprocal teacher-student comprehension, while GSI was marked by unidirectional teacher comprehension, requiring students to allocate greater cognitive resources to process GAI-generated content and engage in self-regulated learning. Moreover, teacher-student brain synchrony was positively associated with students’ social engagement, underscoring its role as a neural signature of effective teacher-student interaction. These findings provide robust neuroscience-based evidence regarding the distinct mechanisms underlying teacher-student interaction and human-AI collaboration, thereby offering theoretical insights and practical implications for optimising instructional design and advancing interdisciplinary research in intelligent educational contexts.