Benchmarking XAI Explanations with Human-Aligned Evaluations

Human judgment should guide XAI evaluation: The PASTA framework demonstrates that explainable AI methods need to be assessed based on how well humans actually understand them, not just technical metrics. The researchers created a large-scale dataset where human evaluators judged different explanation methods, establishing that human perception should be the standard for measuring whether an explanation truly works.

Different explanation types need direct comparison: The research addresses a previously unexamined gap by comparing saliency-based explanations (like highlighting important image regions) against concept-based explanations (like identifying learned features) on equal footing. This cross-modal comparison capability helps organizations choose the most understandable explanation method for their specific needs.

Automated scoring can scale human-aligned evaluation: Rather than requiring human evaluation every time, PASTA-score provides an automated method trained on human preferences that can reliably predict whether explanations will be understandable to people. This makes it practical for organizations to evaluate XAI methods at scale without constant human involvement.

Interpretability assessment can improve AI model design: By using their benchmark to probe how interpretable existing models are, the researchers show that evaluation frameworks can inform the development of inherently more understandable AI systems—moving beyond just explaining opaque models to building more transparent ones from the start.