Algorithmic Bias Detection and Mitigation in AI-Powered Learning Systems Using Machine Learning

AI-powered learning systems, such as intelligent tutoring systems and recommendation engines, enhance education by personalizing learning experiences.However, biases in these systems can lead to inequitable outcomes across gender, ethnicity, or socio-economic groups, exacerbating educational disparities.This study investigates the sources of algorithmic bias, employs machine learning to detect it, and implements mitigation strategies to promote equitable education.An empirical approach was applied, utilizing the Open University Learning Analytics Dataset (OULAD) and synthetic data generated via Conditional Tabular GAN (CTGAN) to ensure balanced demographic representation.The methodology integrates data preprocessing (e.g., SMOTE), fairness-aware algorithms (e.g., adversarial training), and post-processing (e.g., threshold adjustment).Findings indicate that biases stem from skewed training data, algorithmic assumptions, and societal inequities, often resulting in unfair predictions, such as lower pass rates for minority students.The proposed framework reduced Demographic Parity Difference (DPD) from 0.22 to 0.07 and Disparate Impact (DI) from 0.75 to 0.92, with a minimal 2.5% accuracy reduction (88.5% to 86%).Compared to benchmarks like AI Fairness 360's adversarial debasing, this approach achieved superior fairness metrics.Limitations include dataset-specific results and challenges with dataset drift.The framework offers a scalable solution for equitable AI in education, applicable to fair grading and course recommendations, with future research needed for cross-cultural validation and real-time bias correction.