ETD

• algorithmic transparency
• automated decision-making
• explainability by design
• explainable ai
• interpretable machine learning
• right to explanation

This thesis examines automated decision-making (ADM) systems in the administration of justice and argues that post-hoc explainable AI (XAI) techniques, such as LIME and SHAP, cannot satisfy legal transparency demands because they approximate rather than faithfully represent a model's actual reasoning. Drawing on Rudin's critique of "black box" models, it distinguishes complexity-based and proprietary opacity, mapping a spectrum from undisclosed and black-box systems (illustrated by COMPAS) through post-hoc explained models to genuinely interpretable ones, such as CORELS rule lists and linear or logistic regression. Using the Catalan RisCanvi tool as an empirical bridge, it shows how a shift from transparent scoring to undocumented models degrades contestability. The thesis then proposes "explainability by design" as a legal standard grounded in GDPR Articles 22 and 25 and the EU AI Act, built on four criteria: fidelity, contestability, third-party auditability, and proportionality. It argues interpretable-by-design models should be preferred over post-hoc explained black boxes whenever accuracy losses are acceptable, reframing explainability as a threshold requirement for lawful automated adjudication rather than an optional add-on.