Can Explainable AI Explain Unfairness? A Framework for Evaluating Explainable AI

XAI tools can inadvertently enable "fairwashing": While explainable AI tools are designed to make complex models understandable, they can mislead users into trusting biased or incorrect models, creating a false sense of confidence in unfair systems.

Popular XAI tools have significant blind spots for bias detection: The paper found that many prominent XAI tools, despite their strength in simplifying model behavior, lack critical features needed to actually detect bias and fairness issues in the first place.

A framework now exists to evaluate XAI tools for fairness: The researchers developed an evaluation framework that assesses whether XAI tools can detect bias, address fairness concerns, and communicate these findings clearly to users—providing developers with concrete guidance for improvement.

Transparency alone is not enough for responsible AI: Simply explaining how a model works does not guarantee it will reveal or address unfairness; XAI tools need specific design features and capabilities focused on fairness to be truly responsible.

Developers need actionable guidance to reduce fairwashing: The framework suggests specific modifications that XAI toolkit developers can implement to reduce fairwashing risks and ensure their tools actually help identify and address bias rather than obscure it.