ExplainabilityTransparencyTrustworthinesscs.CV
Meta-evaluating stability measures: MAX-Senstivity & AVG-Sensitivity
Miquel Miró-Nicolau, Antoni Jaume-i-Capó, Gabriel Moyà-Alcover
December 14, 2024
xAI
4 citations
Key Takeaways
-
Current XAI evaluation metrics are unreliable: The study found that two commonly used stability measures (AVG-Sensitivity and MAX-Sensitivity) failed to identify random explanations as problematic, suggesting these metrics cannot be trusted to properly assess whether AI explanations are actually trustworthy.
-
The problem of evaluating the evaluators: Rather than just testing AI systems themselves, this research highlights the need to "meta-evaluate"—or check the validity of—the measurement tools used to assess explainability. This is a critical gap because flawed metrics can give false confidence in unreliable AI explanations.
-
Random explanations passed reliability tests: The metrics were unable to distinguish between legitimate explanations generated by a Decision Tree model and completely random explanations and predictions, indicating a fundamental weakness in how we currently measure explanation stability.
-
Implications for responsible AI deployment: If the tools meant to verify that AI explanations are stable and reliable are themselves unreliable, this undermines efforts to ensure AI systems are transparent and trustworthy—a key concern for responsible AI implementation in high-stakes applications.
Abstract
The use of eXplainable Artificial Intelligence (XAI) systems has introduced a set of challenges that need resolution. The XAI robustness, or stability, has been one of the goals of the community from its beginning. Multiple authors have proposed evaluating this feature using objective evaluation measures. Nonetheless, many questions remain. With this work, we propose a novel approach to meta-evaluate these metrics, i.e. analyze the correctness of the evaluators. We propose two new tests that allowed us to evaluate two different stability measures: AVG-Sensitiviy and MAX-Senstivity. We tested their reliability in the presence of perfect and robust explanations, generated with a Decision Tree; as well as completely random explanations and prediction. The metrics results showed their incapacity of identify as erroneous the random explanations, highlighting their overall unreliability.