Recent advances in large language models and immersive interfaces have led to a rapid increase in artificial-intelligence-based instructional avatars. Studies consistently show that these avatars improve perceived social presence and learner engagement. In contrast, the effects on measurable learning outcomes are less consistent. When learning gains occur, they often depend on relational and motivational factors. Therefore, the key question is no longer whether teacher avatars "work," but under which design conditions they can produce lasting educational benefits without causing negative effects such as relational dependency, cognitive overload, mismatched expectations or ethical uncertainty about the teacher's role. The Alma Project argues that research should prioritize pedagogical sustainability rather than maximizing technology. It identifies three design constraints that can preserve the educational relationship while limiting unintended outcomes: using a tutor-facilitator role, adopting minimal embodiment and ensuring controlled and transparent interaction. Building on these principles, we introduce the ALMA (Artificial Learning Mentor Avatar) model. ALMA is a deliberately limited educational avatar designed to provide sustainable instructional support. It is not intended to replace pedagogical authority and remains clearly anchored to the human teacher.
Passalacqua et al. (Thu,) studied this question.