Abstract: The metaverse and artificial intelligence (AI) are increasingly intersecting in educational contexts, yet limited empirical research has examined how generative AI avatars function within socially interactive virtual environments. This study investigates the deployment of generative AI avatars within a socio-educational metaverse environment. Using a developmental evaluation approach, data were collected through interviews with seven institutional stakeholders, teacher-generated reflections, internal documentation, embedded user feedback captured through in-platform reporting tools, and longitudinal field memos across an iterative deployment cycle. Findings indicate that the transition from scripted NPCs to generative AI avatars recalibrated users’ attribution of agency, intensified dialogic unpredictability, and elevated social realism beyond visual fidelity. Voice-mediated interaction emerged as a threshold mechanism for co-presence, while algorithmic improvisation exposed tensions between pedagogical intent and stochastic response generation. The deployment further revealed affective frictions, expectation misalignments, and the mediating role of AI literacy in shaping trust, participation, and interpretive coherence. Overall, the study advances a sociotechnical understanding of AI avatars as co-constructors of meaning and interaction, offering implications for the design, implementation, and governance of future AI-enhanced metaverse learning environments.