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

Researchers present a unified framework addressing a critical gap between algorithmic fairness and explainable AI (XAI): models can produce fair outputs while employing biased reasoning processes. The study introduces the concept of 'procedural bias' and proposes a conditional invariance framework to formalize and audit explanation fairness, establishing the first comprehensive taxonomy and evaluation workflow for this emerging field.

The intersection of algorithmic fairness and explainability reveals a sophisticated vulnerability in AI systems deployed in high-stakes domains. While the machine learning community has developed robust fairness metrics for model outputs and separate XAI techniques for interpretability, this research exposes a blind spot: a system can simultaneously satisfy all standard fairness criteria in its decisions while exhibiting profound unfairness in how it reasons about those decisions. This procedural bias creates a legitimacy problem where stakeholders receive fair outcomes through potentially discriminatory logic.

The research addresses why post-hoc explanation methods fail to certify fairness independently. Post-hoc explainers analyze model behavior after training without access to the underlying decision-making architecture, making them inherently unable to guarantee that explanations themselves are equitable. The conditional invariance framework proposed here formalizes explanation fairness mathematically, requiring that explanations remain invariant with respect to protected attributes when controlling for task-relevant features. This principle subsumes existing explanation fairness metrics as partial implementations.

For AI practitioners and organizations deploying systems in criminal justice, healthcare, credit, and employment, this framework provides actionable infrastructure. The three identified mechanisms generating explanation inequity—representation-driven bias, explanation-model mismatch, and actionability-driven bias—offer diagnostic tools for auditing systems. The six-step evaluation workflow translates theoretical concepts into practical audit procedures, enabling organizations to validate not just fair outcomes but fair reasoning processes.

The work signals a maturation of AI governance discourse, moving beyond surface-level fairness metrics toward deeper procedural transparency. As regulatory pressure intensifies globally, explanation fairness audits will likely become compliance requirements, making this framework increasingly relevant to stakeholders across industries.

• →Models can produce fair outputs while using biased reasoning, creating a 'procedural bias' gap between fairness and explainability.
• →Post-hoc explanation methods cannot certify fairness because they lack access to underlying decision architectures.
• →The conditional invariance framework provides a mathematical foundation unifying existing explanation fairness metrics.
• →Three mechanisms drive explanation inequity: representation-driven bias, explanation-model mismatch, and actionability-driven bias.
• →A practical six-step evaluation workflow enables organizations to audit explanation fairness in production systems.