Fairness of Explanations in Artificial Intelligence (AI): A Unifying Framework, Axioms, and Future Direction toward Responsible AI

Fair outcomes don't guarantee fair reasoning: The paper identifies a critical blind spot where AI systems can produce equitable decisions while using biased reasoning processes—a problem called "procedural bias." This means that checking whether an AI's final decisions are fair is insufficient; we must also examine how the AI explains its reasoning.

Explanations themselves can discriminate based on protected attributes: The research demonstrates that AI explanations can treat people differently based on characteristics like race, gender, or age, even when the actual decisions appear fair. For example, two equally qualified job applicants might receive different explanations for why they were rejected, with one group receiving more detailed or favorable reasoning than another.

Post-hoc explanations have fundamental limitations: The paper warns that adding explanations to AI systems after they're built (a common practice) cannot reliably ensure fair explanations. This suggests organizations need to consider fairness of reasoning from the design stage, not as an afterthought.

A unified framework is needed for explanation fairness: The authors propose that explanations should be "indifferent" to protected attributes—meaning the same relevant factors should generate similar explanations regardless of someone's race, gender, or other protected characteristics. This provides a concrete principle for evaluating whether explanations are truly fair.

Responsible AI requires attention to both outcomes and processes: For high-stakes decisions in criminal justice, healthcare, credit, and employment, fairness means ensuring not just equitable results, but also equitable and transparent reasoning that doesn't discriminate in how it justifies decisions to different groups.