Explainable AI Systems Must Be Contestable: Here's How to Make It Happen

Contestability in AI is currently undefined and unenforceable: Despite being mandated by AI regulations worldwide, there is no formal definition of "contestability" (the ability to challenge AI decisions), no algorithm that guarantees it, and practitioners lack concrete guidance on how to implement it—making current regulatory requirements difficult to satisfy in practice.

A comprehensive framework now exists to make contestability real: The paper introduces the first rigorous formal definition of contestability along with a modular framework spanning human-centered interfaces, technical architectures, legal processes, and organizational workflows, converting what has been an empty regulatory promise into actionable guidance.

Measurement tools can now assess contestability: The authors developed the Contestability Assessment Scale, a composite metric based on over twenty quantitative criteria, enabling organizations to evaluate whether their AI systems actually provide meaningful opportunities for users to challenge decisions—moving beyond vague compliance claims.

Most state-of-the-art AI systems currently fail to be truly contestable: Case studies across diverse application domains reveal that existing systems fall short of genuine contestability standards, highlighting a significant gap between regulatory intent and current practice that organizations need to address.

Contestability requires multiple layers, not just technical fixes: Effective contestability demands coordination across human-centered design, technical architecture, legal processes, and organizational workflows—meaning no single technical solution alone can ensure accountability; organizations must embed contestability into their entire decision-making systems.