Explainability in Artificial Intelligence: Giving a Method to the Madness

Abstract

Explainable AI (XAI) has emerged as a critical response to the increasing opacity of advanced machine learning systems, particularly as their predictive capabilities grow while comprehensibility diminishes. This comprehensive article examines the evolution of explainability methods across three primary categories: model-agnostic approaches that function independently of underlying architectures, model-specific techniques that leverage internal structural knowledge, and inherently interpretable systems designed with transparency as a foundational principle. The article evaluates these methodologies against essential criteria, including fidelity, stability, user comprehensibility, and domain appropriateness, with special focus on highly regulated sectors where explanations are not optional but legally required. The article goes further into the new frontiers of investigation,  including counterfactual expositions, causal interpretability frameworks, and the combination of explainability with fairness aspects. The field has made considerable advances, but it still struggles to standardize measurement of evaluation, deal with vulnerability to adversarial manipulation, and reconcile technical explanations with human cognitive patterns, which indicates a direction towards finding the middle ground between mathematical correctness and practical access to a wide range of stakeholders.