From Flexibility to Manipulation: The Slippery Slope of XAI Evaluation

Explainability evaluation lacks objective standards: Because there is no "ground truth" for what makes a good explanation, researchers must make subjective choices about how to evaluate AI explanations, creating vulnerability to bias and manipulation.

Small technical choices can dramatically change results: The paper demonstrates that seemingly minor adjustments to evaluation settings (hyperparameters) can significantly alter which explanation methods appear to work best, meaning published findings may depend more on researcher choices than actual performance.

Current practices enable intentional or unintentional bias: The flexibility in evaluation methods means researchers could—deliberately or unknowingly—adjust settings to make their preferred explanation approach look better, undermining the credibility of XAI research and real-world deployment decisions.

Transparency and consistency are critical safeguards: The authors propose testing explanation methods across multiple evaluation settings rather than relying on a single configuration, which would make it harder to manipulate results and easier to spot when findings are unreliable.

Responsible AI requires rethinking how we validate explanations: This work highlights that building trustworthy AI systems requires not just better explanation techniques, but more rigorous, standardized, and transparent methods for evaluating whether those explanations actually work.