The Personalization Paradox: Trade-offs Between Social Presence and Task Efficiency in Embodied AR Instructors

Large language models continue to become utilized in training situations to power embodied virtual instructors in Mixed Reality (MR). As these models increase in sophistication, a key question emerges: does designing an agent with similarity to the instructee improve outcomes? We present a user study with four guided assembly conditions: a non-matched instructor employing real-life instructor’s attributes, a personality-matched instructor, a gender- and voice-matched instructor, and a fully matched instructor reflecting the user’s Big Five personality, cloned voice, and gender. Participants completed an ordered assembly task and reported on instructional quality. Results show that fully matched instructors were overwhelmingly preferred and significantly enhanced social presence and user experience. However, these subjective benefits did not translate into faster task completion, revealing a trade-off between engagement and efficiency. These findings offer critical guidance for designing future embodied virtual instructors and highlight the nuanced role of personalization in human–AI interaction.