Automated Bias Assessment in AI-Generated Educational Content Using CEAT Framework

TL;DR

The paper tackles bias in AI-generated educational content by developing an automated bias auditing pipeline that combines Contextualized Embedding Association Test (CEAT) with prompt-engineered word extraction in a Retrieval-Augmented Generation (RAG) framework. It demonstrates that automated word extractions closely align with manually curated ground-truth sets, achieving a near-perfect Pearson correlation between automated and ground-truth CEAT scores ($r = 0.9930$) and strong semantic alignment (cosine similarities $0.7627$–$0.8895$). This approach reduces subjectivity and scales bias assessment for educational materials, supporting fairness and reproducibility in GenAI-driven tutoring content. The work discusses educational implications, acknowledges limitations such as dataset size and lack of mitigation, and points to future directions in bias mitigation and broader classroom validation. Overall, the method provides a scalable, objective tool for auditing demographic biases in AI-generated educational content with potential policy and practice impact.

Abstract

Recent advances in Generative Artificial Intelligence (GenAI) have transformed educational content creation, particularly in developing tutor training materials. However, biases embedded in AI-generated content--such as gender, racial, or national stereotypes--raise significant ethical and educational concerns. Despite the growing use of GenAI, systematic methods for detecting and evaluating such biases in educational materials remain limited. This study proposes an automated bias assessment approach that integrates the Contextualized Embedding Association Test with a prompt-engineered word extraction method within a Retrieval-Augmented Generation framework. We applied this method to AI-generated texts used in tutor training lessons. Results show a high alignment between the automated and manually curated word sets, with a Pearson correlation coefficient of r = 0.993, indicating reliable and consistent bias assessment. Our method reduces human subjectivity and enhances fairness, scalability, and reproducibility in auditing GenAI-produced educational content.