Explainable Artificial Intelligence: a Systematic Review

The "accuracy vs. explainability" trade-off is a central challenge in AI: While modern machine learning models, especially deep learning systems, have become increasingly accurate, this improvement has come at the cost of reduced explainability and interpretability—meaning it's often unclear how these powerful systems actually make their decisions.

A diverse toolkit of XAI methods now exists: The field has developed numerous approaches to make AI systems more transparent, which the review organizes into four main categories: review articles, theoretical foundations, practical methods, and evaluation techniques—indicating that solutions are available but practitioners must choose appropriately for their context.

Explainability is not one-size-fits-all: The systematic review's hierarchical classification system suggests that different XAI methods serve different purposes and audiences, implying that organizations deploying AI need to carefully consider which explanation approach best serves their specific stakeholders (regulators, end-users, developers, etc.).

The field still lacks standardized evaluation approaches: By identifying "evaluation" as a distinct cluster requiring attention, the review highlights that there's no consensus on how to measure whether an AI explanation is actually good or useful—a gap that could hinder responsible AI deployment until standards are established.

XAI research is rapidly evolving but fragmented: The significant recent growth in XAI methods suggests both opportunity and confusion; without clear guidance on which approaches work best for different applications, organizations may struggle to implement explainability effectively.