Explanations as Bias Detectors: A Critical Study of Local Post-hoc XAI Methods for Fairness Exploration

Explanation methods can help detect unfairness, but require careful validation: The paper demonstrates that local post-hoc explanation methods (techniques that explain individual AI decisions after the fact) can be useful tools for identifying bias, but their results must be carefully scrutinized rather than automatically trusted as reliable bias detectors.

Different explanation methods may give inconsistent results about fairness: A key finding is that the consistency and quality of fairness insights varies across different explanation methods, meaning organizations cannot rely on a single explanation technique to comprehensively assess whether their AI systems treat protected groups fairly.

Simply removing protected attributes doesn't guarantee fairness: The research addresses a common misconception that removing sensitive information (like race or gender) from AI systems automatically prevents discrimination. The paper shows this approach requires more nuanced consideration of how fairness actually works in practice.

How you measure fairness matters as much as detecting it: The study highlights the distinction between "distributive fairness" (whether outcomes are fair) and "procedural fairness" (whether the decision-making process is fair), showing that explanation methods may reveal different types of bias depending on which definition you prioritize.

Aggregating individual explanations into group-level fairness assessments requires careful methodology: The paper emphasizes that how organizations combine explanations from individual cases significantly impacts conclusions about whether AI systems are fair to entire groups, requiring deliberate choices about aggregation strategies.