AI Fairness and Explainability: A Unified Framework for Procedural Bias
A recent survey paper published on arXiv highlights a significant oversight at the crossroads of algorithmic fairness and explainable AI (XAI). The researchers contend that a model may meet conventional fairness standards in its results yet exhibit substantial unfairness in its reasoning, a concept they refer to as 'procedural bias.' This paper represents the inaugural comprehensive theoretical and literature review in this developing area, addressing critical sectors such as criminal justice, healthcare, credit, and employment. It challenges the effectiveness of post-hoc explainers in ensuring explanation fairness and suggests a set of axioms along with a framework aimed at promoting responsible AI.
Key facts
- The paper is a survey on fairness of explanations in AI.
- It identifies procedural bias as a novel blind spot.
- Procedural bias occurs when a model's reasoning process is unfair despite fair outputs.
- The paper covers high-stakes decisions in criminal justice, healthcare, credit, and employment.
- It provides the first unified theoretical and literature review of this field.
- It critiques post-hoc explainers for certifying explanation fairness.
- The paper proposes axioms and a framework for responsible AI.
- The paper is published on arXiv with ID 2605.09852.
Entities
Institutions
- arXiv