Visual Authenticity Outweighs Vocal Naturalness: An Eye-Tracking Study on Generative AI in Micro-Learning

With the growth of micro-learning on short-form video platforms, instructional videos are increasingly produced using AI-generated voices and avatars. This study examined how voice type (human vs. AI-generated) and avatar type (human vs. AI-generated) influence immediate learning, visual attention, physiological responses, and subjective evaluations in 60-second instructional videos. Thirty participants completed a 2 × 2 within-subject experiment while learning performance, AOI dwell time, blink rate, CFF change, and post-task ratings were recorded. Human-avatar conditions produced higher objective learning performance than AI-avatar conditions, whereas human voices improved perceived comprehension without significantly improving objective scores. The AI voice–human avatar condition showed reduced blink rate and relatively low dwell time on instructional content, suggesting possible increases in attentional control or processing demand. No detectable CFF change was observed. These findings indicate that subjective preference alone is insufficient for evaluating AI-mediated micro-learning and that visual instructor authenticity warrants particular design attention.