New study of 43 cases shows how institutions resist AI accountability

A research paper accepted at FAccT 2026 provides an empirically grounded analysis of how AI contestation actually works in practice. The authors examined 43 real-world cases where affected groups—including workers, communities, and public-interest organizations—directed explicit demands for accountability toward organizations developing or deploying AI systems. Using Bovens's relational model of accountability as a framework, the study conceptualizes contestation as an iterative, dynamic process where actors 'from below' seek redress, influence, or systemic change, while actors 'from above' respond by accepting, resisting, or circumventing those demands.

The analysis produced detailed categories of contestation strategies, institutional response tactics, and outcome types. A key finding is that institutions frequently deploy a range of strategies to limit their accountability, from deflecting blame to offering superficial concessions. The paper provides specific guidance for researchers, policymakers, and advocates to anticipate and counter these evasion tactics, helping ensure that AI accountability mechanisms are more than just performative. This work fills a critical gap in understanding the real-world power dynamics behind AI governance.