Unifying VXAI: A Systematic Review and Framework for the Evaluation of Explainable AI

The XAI field lacks standardized evaluation methods: Despite growing numbers of explainability techniques for AI systems, there is currently no consensus on how to properly evaluate whether these explanations are actually useful or trustworthy—a critical gap the paper addresses by identifying 41 different metric groups from 362 publications.

Not all explanations are equally valuable: The paper establishes that explanations must be rigorously evaluated against specific quality standards to ensure they genuinely help people understand AI decisions, rather than simply appearing to do so, which is essential for building genuine trust in AI systems.

A unified framework is now available for evaluation: The VXAI framework provides organizations and researchers with a structured, comprehensive approach to selecting and comparing explainability methods, making it easier to ensure consistency and comparability when evaluating different AI explanation techniques.

Complexity of AI systems demands transparency solutions: The paper emphasizes that modern deep neural networks with millions of parameters are inherently difficult to understand, making explainability tools not just helpful but necessary for responsible deployment of powerful AI systems in high-stakes applications.

Standardization enables better AI governance: By organizing evaluation metrics into a coherent framework with clear categorization schemes, the paper supports the development of standardized protocols that can help organizations systematically assess and compare XAI methods before deployment.