Explainable AI Isn't Enough! Rethinking Algorithmic Contestability

Explainability alone is insufficient for fairness: The paper demonstrates that popular XAI methods like LIME, counterfactuals, and Anchors—while helpful for understanding decisions—fail to provide adequate evidence for challenging incorrect algorithmic decisions, revealing a critical gap between explanation and accountability.

Contestability deserves equal focus to recourse: While most AI ethics work emphasizes helping people change their features to get better outcomes (recourse), the paper argues that contestability—the ability to prove an algorithmic decision was wrong and overturn it—is equally important but has been largely neglected in research and practice.

Current explanations only work locally: The paper finds that standard XAI explanations can only identify errors in cases similar to an individual's own situation, but provide insufficient grounds to overturn the specific decision affecting that person, limiting their usefulness for challenging high-stakes decisions like loan denials or hiring rejections.

Legal and ethical stakes demand better tools: Because machine learning systems make consequential life decisions about loans, employment, and academic integrity, individuals need formal mechanisms to contest and correct erroneous decisions—not just understand why they were made—making contestability a matter of both fairness and legal compliance.

Three types of evidence are needed: Beyond traditional explanations, the paper identifies that overturning algorithmic decisions requires multiple forms of evidence, though the abstract cuts off before fully detailing these categories.