AI-Generated Avatars in Educational Videos : Assessing how AI Disclosure Timing Affects User Perception and Memory Retention

Sammanfattning

Human-like avatars generated by Artificial Intelligence (AI) are increasingly being explored as tools in educational settings, yet little is known about how disclosure of their AI nature affects user's perception and learning outcomes. This study examined how the timing of AI disclosure influences trustworthiness, perceived humanness, emotional responses, and memory retention in an educational video. A between-subjects experimental design was used with 60 participants (36 female, 24 male) aged 18-40. Participants were randomly assigned to one of three groups differing in the timing and type of AI disclosure: before viewing, after viewing, or deception after viewing with false information about a human presenter being AI-generated. Data was collected through Likert-scale items, open-ended questions, and a 20-item multiple-choice memory test. Quantitative data were analyzed using ANOVA and MANOVA, and qualitative data were analyzed using a content analysis. No statistically significant group differences were found across any of the quantitative variables. However, observed trends in the quantitative and qualitative data suggested that prior disclosure was associated with lower perceived humanness and greater analytical attention toward the presenter, without impairing memory performance. Qualitative data also indicated that expert authorship preserved trustworthiness across groups, while non-disclosure evoked feelings of deception. These findings suggest that realistic AI-generated avatars may support learning outcomes regardless of disclosure timing, though transparency remains important for user trust and ethical AI communications.