Building Trust in AI-Powered Assessment Through Explainable Machine Learning Models

The increasing integration of artificial intelligence (AI)-based assessment systems in higher education offers efficiency and scalability in grading and feedback but raises concerns about transparency, fairness, and student trust. This study investigates the role of Explainable AI (XAI) in improving trust and engagement within AI-driven assessment environments, focusing on the Thuto Learning Management System (LMS) at the National University of Lesotho (NUL). A cross-sectional survey of 850 undergraduate students, who had engaged with Thuto LMS for at least two semesters, was conducted using a 55-item questionnaire measuring 11 constructs, including digital self-efficacy, AI literacy, ethics awareness, and feedback perceptions. Assessment outcomes were retrieved from LMS records and binarised into pass/fail classifications. Three supervised machine learning models—Logistic Regression, Random Forest, and K-Nearest Neighbours—were developed to predict assessment outcomes, and post hoc interpretability was achieved using SHAP and DALEX frameworks. Logistic Regression demonstrated the highest predictive accuracy (73.7%), while feature importance analyses revealed that feedback usefulness, discussion engagement, and digital self-efficacy were the strongest predictors of academic success. Findings underscore the potential of XAI for promoting fairness, transparency, and learner trust in AI-powered educational systems, particularly in underexplored African higher education contexts.