Explainable Artificial Intelligence: A Review of Techniques, Applications and Challenges with a Proposed Framework for Comparing LIME and SHAP

Artificial Intelligence (AI) models such as ensemble methods and deep neural networks achieve high accuracy but behave as black boxes, which limits trust, auditing and regulatory acceptance. Explainable Artificial Intelligence (XAI) aims to make model decisions understandable to humans. This paper presents a structured review of XAI covering core concepts (interpretable versus explainable models, local versus global explanations, intrinsic versus post-hoc methods), the working principles, strengths and limitations of two widely used model-agnostic techniques, LIME and SHAP, and of feature importance, partial dependence plots, counterfactual explanations, Grad-CAM and Integrated Gradients. Published studies from 2018 to 2026 on the reliability of explanations, comparative evaluations and applications in healthcare, finance, cybersecurity, education, autonomous vehicles and industry are synthesised, and a research gap is identified: explanation methods are rarely compared with a common, multi-metric protocol. To address this, a proposed (not yet implemented) framework for comparing LIME and SHAP using fidelity, stability, interpretability, computational efficiency and robustness is described, together with a literature-based discussion of expected outcomes. This is a review article; no new experimental results are reported.