The increasingly sophisticated anthropomorphism of AI-powered pedagogical agent design raises critical questions about how human-like features shape learning dynamics. Addressing the underexplored interaction between human-like features and stereotype activation in STEM education, this study investigates pedagogical implications through the Stereotype Content Model while focusing on the uncanny valley effect as a critical moderator. Combining experimental manipulation with mixed methodologies, the study critically engages with the paradoxical relationship between technological realism and educational effectiveness in STEM learning environments. Through systematic manipulation of appearance (2D vs. 3D), clothing (casual vs. formal coats), and gestural dynamics, pedagogical agents with distinct anthropomorphism levels (low/moderate/high) were selected to unravel how perceived warmth and competence mediate learner acceptance and performance (N = 410 learners). The analysis reveals three critical insights: First, anthropomorphism operates as a dual-channel mechanism that differentially activates learners' social cognition, with competence demonstrating a stronger mediation effect on learning performance compared to perceived warmth (43.3% vs. 28.57%), likely due to STEM learners' prioritization of task-relevant cues; Second, the uncanniness manifests as a non-linear moderator, suggesting threshold effects in human-computer interaction design; Third, no direct effect of anthropomorphism on performance was identified, underscoring the necessity for competence-driven design strategies. These findings highlight the threshold effects of the uncanny valley in pedagogical agent design.
Hu et al. (Mon,) studied this question.