Abstract

Machine learning now informs consequential decisions in education. It flags students at risk, forecasts attainment, recommends interventions. The models that predict best are the least transparent, and this opacity now threatens trust, accountability, and the pedagogical legitimacy of these systems. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) are the post-hoc techniques most often proposed as the remedy. How the two are actually used in educational research is far less clear. This review, reported in accordance with the PRISMA 2020 statement, covers 75 peer-reviewed studies published between 2022 and 2026. Five databases were searched (Scopus, SpringerLink, ScienceDirect, ERIC, and IEEE Xplore) and returned 781 records; 669 remained after de-duplication, 135 met the eligibility criteria, and 75 full texts were retrieved, coded, and analysed by model, dataset, task, and explanation method. The picture that comes out of the coding is one-sided. SHAP appears in 74 of the 75 studies. LIME appears in 33, and in 32 of those 33 it accompanies SHAP rather than replacing it. Tree-based models predominate (72 of 75), and the literature concentrates on performance and dropout prediction in higher education. Interpretability is judged almost entirely from the model side; whether educators or students understand the explanations, or change anything because of them, is rarely tested. The review describes this state of practice as SHAP-centric, technically framed interpretability, and uses the gaps it leaves, in human-centred evaluation, longitudinal explanation and fairness, to set out a research agenda for trustworthy educational XAI.