Designing Explainable and Trustworthy AI Decision Support for Resource-Constrained Courts: Evidence from Kenya's Magistrate Courts

Courts are adopting artificial intelligence faster than they are learning how to govern it. The stakes of that gap are highest where judicial institutions are least resourced, data infrastructure is weakest, and no regulatory framework for judicial AI yet exists. This paper examines how an explainable, ethical, and operationally trustworthy AI decision-support system can be designed, validated, and deployed inside such an institution, using the Judiciary of Kenya’s Integrated Case Management System (ICMS) as the empirical setting. Drawing on 672,974 validated magistrate court case records, the paper presents a threestage prediction pipeline, a same-day triage classifier, a filing-time duration model, and a dynamic progression model, not as a modelling exercise but as a set of institutional design choices: staged prediction matching the case lifecycle, SHapley Additive exPlanations (SHAP) integrated from the outset so that every prediction is inspectable by non-technical staff, deliberate exclusion of outcome data to prevent leakage, an explicit boundary against magistrate performance evaluation, and a containerised, logged, token-authenticated deployment architecture built for auditability. The system’s transparency surface is analysed for three user groups, administrators, magistrates, and judicial leadership, and its ethical exposure is assessed across magistrate profiling, geographic disparity amplification, litigant transparency, accountability, and emerging AI governance frameworks including the EU AI Act and the African Union’s continental AI strategy. The Kenyan experience yields transferable conditions under which judicial AI in developing-world courts can earn, rather than assume, institutional trust.