Proceedings of the Annual Meeting of the Cognitive Science Society

Text-to-image (T2I) models are increasingly used to generate educational visuals, from textbook illustrations and lecture slides to classroom scenes and institutional marketing materials. However, these models do not neutrally depict education. They often encode stereotyped views of what education should look like, and such patterns can recur across generations and gradually solidify into taken-for-granted visual common sense. Building on cognitive theories, we introduce the notion of Educational Epistemic Bias to describe how T2I models narrow rich educational ideas into a small set of recurring visual patterns. We define this construct along four dimensions and design an education-specific benchmark of 60 prompts that span learning environments, activities, power relations, and roles. We apply this benchmark to nine mainstream T2I models, yielding 2,160 images. To quantify bias, we propose the Educational Epistemic Biases Quantifier (EEBQ), a VLM-based evaluation framework that includes two image-quality metrics and four metrics targeting DEI-related aspects of the images. Our analysis reveals systematic educational epistemic biases across all models. Together, the benchmark and EEBQ offer a concrete way to examine how generative models visually construct education and to inform more careful use of such systems in educational settings.