The integration of generative Artificial Intelligence (AI) models in educational content creation presents unprecedented opportunities alongside significant challenges related to bias propagation and stereotype reinforcement. Current language and image generation models, trained on vast internet-sourced datasets, inherently carry biases reflecting predominantly first-world perspectives and various cultural, gender, and societal stereotypes. When deployed in educational contexts, particularly in dynamic, interactive environments, these models risk perpetuating harmful biases that compromise educational objectives and competency development. This paper presents an innovative multi-agent architecture designed to address these challenges through automated self-reflection mechanisms. The system comprises two primary AI agents: a generator agent utilizing state-of-the-art language or image generation models to create initial educational content, and a critic agent employing either language models or vision-language models to evaluate generated materials against educator-defined ethical and pedagogical criteria. The proposed architecture has been successfully tested and evaluated across diverse educational challenges, demonstrating its effectiveness in producing culturally appropriate, age-suitable content while avoiding conflicts and biases inherent in standard AI models.

Keywords: Generative Artificial Intelligence, Education, Content Creation, Technology, Ethics.