What will it take to generate fairness-preserving explanations?

Explanations of AI systems don't automatically inherit fairness from the model itself. Even if a black-box algorithm has been designed or audited to be fair, the explanations generated to describe how it works may not preserve or accurately represent those fairness properties, potentially misleading users about how fairly the system actually operates.

Explanation algorithms can obscure critical fairness-relevant information. The paper identifies that current explanation methods may ignore or hide important factors related to fairness, meaning explanations could present an incomplete or distorted picture of whether a model treats different groups equitably.

There's currently no standard way to evaluate whether explanations are "fairness-preserving." The research highlights a gap in how we assess explanations—we lack established methods to determine if an explanation accurately reflects both what a model does AND whether it does so fairly, which is essential for responsible AI deployment.

The problem is particularly acute for tabular datasets (structured data with rows and columns). While the paper focuses on this common data format used in high-stakes decisions like lending and hiring, it suggests the fairness-explanation gap likely extends to other domains where explanations are used to justify consequential decisions.

New research is needed to develop better explanation methods that prioritize fairness. The authors call for future work specifically designed to create explanations that are both informative about model behavior AND transparent about fairness implications—not just one or the other.