Life of Research · open-access repository · LOR-00208
Faithful, Robust, Uncertainty-Aware and Human-Centred Explanations for High-Stakes Machine Learning Systems
- University of Cape Town
Thesis / Dissertation
LOR-00208
Permanent link: https://lifeofresearch.com/theses/208
Abstract
Machine learning systems now inform decisions in medicine, finance, justice and public administration, yet the most accurate models are frequently opaque, brittle under distribution shift, and unevenly reliable across population groups. Explainable AI (XAI) promises to close the gap between predictive performance and warranted human trust, but a growing body of evidence shows that popular explanation methods can be unfaithful to the model, unstable under small perturbations, manipulable by adversaries, and of limited help to the people who must act on them. This thesis argues that explainability and trustworthiness must be studied as a single, jointly evaluated problem rather than as separate add-ons to a trained model.
The thesis makes four contributions. First, it proposes a conceptual framework that links explanation properties (fidelity, stability, comprehensibility, actionability) to trustworthiness pillars (robustness, calibrated uncertainty, fairness, accountability) and to stakeholder needs. Second, it develops a methodology for training and evaluating faithful and stable feature attributions. Third, it introduces uncertainty-aware explanations that report how confident the model, and the explanation itself, are. Fourth, it designs a human-centred evaluation protocol, including fairness-oriented audits and controlled user studies, to measure whether explanations improve appropriate reliance rather than merely increasing confidence. The thesis closes with an integrated assessment toolkit and a discussion of regulatory alignment with the EU AI Act and the NIST AI Risk Management Framework.