Explainable AI (XAI): Methods, Challenges, and Future Directions – A Systematic Review

Abstract

Abstract Explainable artificial intelligence (XAI) is becoming a multidisciplinary and crucial research area aimed at resolving the opacity and interpretability issues of stakeholders in high-stakes applications of complex machine learning models. AI increasingly shapes decisions in healthcare, finance, autonomous systems, and critical infrastructure sectors, where transparency, accountability, and trustworthiness are paramount. This systematic literature review synthesizes current research on XAI by following the PRISMA guidelines to identify the main approaches, assess enduring challenges, and suggest future avenues of research. A total of 98 papers were obtained from Google Scholar using a detailed search strategy, and these underwent full-text evaluation through rigorous screening processes, providing 57 papers for full-text analysis. This survey organizes XAI methods into pre-modeling interpretable and post-hoc explanation methods, including LIME, SHAP, Grad-CAM, attention mechanism, counterfactual explanation, and concept-based methods. Some challenges identified include the basic trade-off between model performance and explainability, the absence of standard evaluation frameworks, the fragility of explanations against adversarial perturbations, and the conflict between the faithfulness and plausibility of explanations. Application domains include healthcare, Industry 4.0, energy systems, and cybersecurity, which have unique interpretability requirements. Future research should focus on unifying evaluation standards for complex XAI models via a combination of automated fidelity tests and human-centered validation, formalized causal/counterfactual explanation methods, integration of robustness-testing protocols in comprehensive XAI pipelines, and domain standards for regulated industries. This review offers a systematic overview of the state-of-the-art, exposing both recent methodological contributions that aim to address its many challenges and major gaps that still need to be filled with respect to the use of trustworthy AI in real-world applications.