Interpretability and Explainability in Modern AI Systems: Concepts, Methods, and Challenges

Abstract

Explainable Artificial Intelligence (XAI) has become essential for ensuring transparency, accountability, and trust in modern AI systems, particularly as machine learning and deep learning models grow increasingly complex. The survey is the in-depth analysis of the foundations of the XAI, its key definitions, the concepts of interpretability, and the paradigms of global and local explanations. It discusses the intrinsic and post-hoc methods of interpretation, including model-specific interpretability, feature attribution methods, including SHAP and LIME, and visualization-based explanation approaches. The survey also discusses the explainability challenges in deep neural networks, transformers, and generative models, and along with emerging approaches such as concept-based explainability and neurosymbolic explainability. The urgent need for interpretable AI in high-stakes domains is highlighted in applications, healthcare, finance, autonomous vehicles, cybersecurity, and public policy. Although significant progress has been achieved, there are still issues with evaluating metrics, balancing explainability and privacy and security, and scaling XAI to large models. The paper outlines some promising areas of future research, such as interactive XAI, AutoXAI, causal interpretability, and XAI of large language models. Altogether, this survey highlights the growing importance of XAI in the developing of trustful, ethical, and human- centered AI systems.