How Embodiment Changes AI Tutoring: Comparing Student Perceptions of Text and AI Human Avatar-Based Chatbot Systems

Embodied conversational interfaces are able to deliver generative AI tutoring via synthesised speech and human-like video avatars. However, there is limited empirical evidence on how embodiment shapes the learner experience and self-regulated learning (SRL). This study compares the perceptions of students interacting with a text-based Retrieval-Augmented Generation (RAG) tutor and an avatar-based RAG tutor in two authentic higher education contexts (introductory computer science and Python programming) across two European universities. Adopting a Design-Based Research approach, we collected (1) questionnaire data from users and non-users, and (2) anonymised interaction logs comprising 341 user messages, which were coded using an SRL process–action framework. The results show that the perceived level of learning support is largely consistent with previous text-only deployments: the tutor is primarily valued for providing explanations and support related to practice, while motivational support is still comparatively weak. However, embodiment notably improves perceptions of the system as a communication partner, with active users rating this dimension more positively than non-users. Modality ratings suggest that audio output is perceived as more helpful than the visual avatar. Log analyses show that interactions are still dominated by information-seeking (Seeking–Search) and that metacognitive SRL actions (e.g. goal setting and self-evaluation) are rare. Embodiment also introduces additional peripheral interactions (e.g. character-focused prompts and human-likeness checks), thereby increasing the proportion of non-learning dialogue. Overall, the findings suggest that, while embodiment enhances social presence, it does not automatically foster deeper SRL. Explicit pedagogical scaffolding is needed to encourage learners to engage in metacognitive regulation.