Focus! Rating XAI Methods and Finding Biases

Most popular AI explanation methods are not equally reliable: The paper evaluated six widely-used explainability techniques and found significant differences in their quality. Layer-wise Relevance Propagation (LRP) and GradCAM consistently produced trustworthy explanations, while SmoothGrad and Integrated Gradients produced explanations that didn't properly distinguish between different classes, potentially misleading users about why models make decisions.

Current methods for testing AI explanations have a major flaw: Existing approaches typically test explanations by adding artificial noise to images, which doesn't reflect real-world conditions. This paper introduces a more realistic testing method using image mosaics from different classes, providing a better way to verify whether explanations actually focus on relevant features.

Explanation quality can reveal hidden model biases: Beyond just rating explanation methods, the paper demonstrates that Focus can be used as a tool to identify and characterize biases in AI models themselves—meaning better explanation evaluation methods can help catch problematic patterns that models have learned.

We need standardized metrics for AI transparency: The paper addresses a critical gap in responsible AI: there were no objective standards for assessing explainability methods before this work. Establishing consistent evaluation scores like Focus is essential for ensuring that when companies claim their AI is "explainable," that claim is actually meaningful and verifiable.