Explainable Artificial Intelligence (XAI) Frameworks for Transparent Decision-Making in High-Stakes AI Applications: A Comprehensive Review

Abstract

ABSTRACT The rapid proliferation and operational deployment of black-box Artificial Intelligence (AI) models—specifically deep neural networks, ensemble architectures, and transformer variants—in high-stakes domains such as clinical medicine, algorithmic finance, autonomous driving, and legal adjudications have catalyzed urgent concerns regarding transparency, accountability, safety, and algorithmic bias. Explainable Artificial Intelligence (XAI) has emerged as an indispensable paradigm to bridge the critical gap between predictive performance and human interpretability. This review paper provides an exhaustive, highly technical analysis of modern XAI methodologies, evaluation frameworks, and operational bottlenecks across critical application domains. We propose a comprehensive taxonomy classifying post-hoc attribution methods (LIME, SHAP, Integrated Gradients, Grad-CAM) alongside intrinsically interpretable glass-box models (Generalized Additive Models, Symbolically Constrained Decision Trees). We rigorously evaluate these frameworks across quantitative dimensions including explanation fidelity, computational complexity, adversarial robustness, and cognitive utility for domain experts. Furthermore, this review systematically analyzes key industry case studies, detailing how XAI algorithms mitigate diagnostic risks, ensure regulatory compliance under GDPR and the EU AI Act, and minimize catastrophic failures in autonomous control. Finally, we highlight fundamental research gaps, including explanation fragility, adversarial manipulation of attribution maps, and the trade-off between post-hoc fidelity and computational latency, outlining concrete strategic roadmaps for future XAI research.

KEYWORDS: Explainable Artificial Intelligence (XAI), Model Interpretability, Post-Hoc Attribution, SHAP, LIME, High-Stakes AI, Algorithmic Transparency, Model Fidelity, Adversarial Robustness.