Navigating the Maze of Explainable AI: A Systematic Approach to Evaluating Methods and Metrics

Different XAI evaluation metrics often contradict each other: The paper found that using different metrics to evaluate explainability methods can lead to conflicting rankings of which methods work best. This means practitioners relying on a single metric risk making poor choices about which explanation tool to use for their specific problem.

No one-size-fits-all explainability method exists: By testing 17 different XAI methods across 7,560 different combinations of model architectures and data types, the research demonstrates that the best explanation method depends heavily on the specific context—including what type of AI model is being used and what kind of data it processes.

Current XAI research has significant blind spots: The study reveals that most existing research only examines a few explanation methods using limited metrics, creating selection bias. Notably, Expected Gradients emerged as a top-performing method despite never being comprehensively evaluated in prior related studies, suggesting important methods may be overlooked.

Practitioners need systematic guidance, not just individual metrics: Rather than relying on single evaluation scores, the paper proposes a more robust evaluation scheme that considers multiple metrics together, helping organizations make more informed decisions when implementing explainability tools in their AI systems.

Transparency in XAI evaluation is critical for responsible AI: By publicly releasing over 326,000 saliency maps and 378,000 metric scores, the research enables the broader community to validate findings and build more trustworthy AI systems grounded in reliable explainability assessments.