Challenges in Applying Explainability Methods to Improve the Fairness of NLP Models

Explainability methods alone don't automatically fix bias: While XAI tools are often promoted as solutions for detecting and reducing bias in NLP models, the paper finds that the connection between "explaining" a model and actually "making it fairer" is frequently unclear and underspecified in research.

Current practices for using explainability to address fairness are limited: The paper identifies that existing approaches applying XAI methods to detect and mitigate bias in NLP are relatively sparse and lack standardized methodologies, suggesting the field hasn't yet developed robust frameworks for translating explanations into fairness improvements.

Significant barriers prevent wider adoption of XAI for fairness work: The research uncovers specific obstacles that stop explainability methods from being more widely used to tackle fairness issues, though the paper indicates these barriers require better understanding and targeted solutions.

The gap between intention and implementation is a key problem: Many XAI methods are motivated by fairness goals, but researchers often fail to specify concrete steps for how explanations would actually help detect, quantify, or mitigate bias in practice—creating a disconnect between theory and application.