AI-Generated Classroom Videos in Education Research: Educator Perceptions of Realism, Engagement Cues, and Usefulness
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Authentic classroom videos are widely used in teacher education and research because they capturecomplex behavioural and social dynamics. However, collecting and sharing such footage is costly andraises concerns about student privacy. Advances in generative AI, especially text-to-video models,now enable fully synthetic classroom videos that offer experimental control without using real studentdata. The key question is whether these videos are realistic and useful enough to serve as alternativesto authentic footage. We report findings from a survey of 128 educators who evaluated 12 short clipsfrom three sources: authentic, staged simulations, and AI-generated videos. Participants rated realism,clarity of engagement and disengagement cues, perceived strengths and limitations, and whether theywould recommend AI-generated clips for research or training. Authentic videos were rated as mostrealistic. AI-generated and staged clips were judged similarly believable but formed a lower realismtier. Educators accurately identified behavioural and emotional cues across all video types, thoughclarity varied by source and engagement state. Reported strengths of AI-generated videos includedexperimental control, efficiency, and privacy. Limitations concerned emotional depth, movementquality, and social complexity. Most respondents viewed AI-generated clips as supplements ratherthan replacements for authentic footage. These findings provide a baseline of educator perceptionsacross realism, ecological validity, social presence, perceptual authenticity, and behavioural fidelity.Based on this, we propose a Zone of Utility for synthetic media. Within this zone, current models aresuitable for controlled, privacy-safe procedural scenarios and for augmenting engagement datasets.Beyond that, limited micro-expressive fidelity makes them unsuitable for training in affective judgmentor for replacing authentic classroom recordings. The tendency to group staged and AI clips into onesynthetic category also suggests a shifting perception of artificial classroom media.