From Code to Character: Investigating Personality in Generative AI-Driven Educational Avatars

This study examines how personality design can shape the experience of generative AI-driven educational avatars, building on Pal et al.'s (2023) eight-trait framework of six positive and two negative personality types. The work centers on a prototype of eight AI chat avatars, each embodying one of these personalities, developed in Unity and powered by Meta's LLaMA 3.1 8B Instruct model. The research is positioned at the intersection of conversational AI, anthropomorphism, and experiential learning theory, drawing on literature in speech synthesis, trust in voice assistants, and robot-mediated education, while situating its design within commercial products such as Khanmigo and Duolingo Max.

The formative evaluation involved 15 engineering students, and its most striking result was the level of engagement. Every group exceeded the planned 15 to 25 minute session, with some sessions extending to 40 minutes, and participants reported heightened motivation and a sense of connection with the avatars. Notably, every group selected the "offensive" avatar, which was valued for its concise responses despite its hostile tone, a finding that complicates assumptions about desirable AI persona traits. Participants perceived the creative avatar as more imaginative and the intelligent avatar as fact-based but less human-like.

The authors conclude that personality should be matched to task demands and user preferences rather than treated as a uniform design goal. They are explicit that these findings are preliminary, given the small, single-discipline sample. Their stated next steps include investigating real-time personality adaptation, cross-cultural perception of AI personas, and the long-term effects of personality design on learning outcomes.