TITLE:
A Systematic Review of Multimodal AI Agents and Embodied Avatars in University EFL Speaking Support
AUTHORS:
Qixin Zhu
KEYWORDS:
Multimodal AI Agents, University EFL Speaking, Embodied Avatars, Social Presence, Cognitive Load
JOURNAL NAME:
Open Journal of Social Sciences,
Vol.14 No.3,
March
17,
2026
ABSTRACT: The integration of Artificial Intelligence into EFL education has historically prioritised text-based accuracy. However, the period between 2024 and 2026 witnessed a multimodal turn characterised by high-fidelity embodied avatars. This systematic review examines the empirical landscape of Multimodal AI Agents (MMAAs) specifically for university-level EFL speaking support. By synthesising data from 82 empirical studies (N = 82) indexed in Web of Science and Scopus, this report identifies critical themes: the paradox of social presence in anxiety regulation, the instructional implications of AI Twins, the cognitive load of the uncanny valley, interactional bottlenecks in turn-taking, the pedagogical shift from tutor to peer, and the ethical risks of biometric surveillance. Findings suggest that while embodied agents significantly enhance Willingness to Communicate (WTC) and emotional engagement, their efficacy depends more on artificial coherence than photorealism. Furthermore, the emergence of AI Peers challenges traditional hierarchies, suggesting a future in which AI functions as a fallible, tireless social partner rather than an omniscient authority. The review concludes that the primary value of MMAAs lies in decoupling social practice from social risk, though this potential is contingent upon balancing affective benefits with cognitive constraints and rigorous ethical safeguards. The review concludes that the primary value of MMAAs lies in decoupling social practice from social risk, provided that educators actively mitigate the interview effect and prioritise artificial coherence over photorealism.