New fairness tests help certify or detect bias within allowed limits
Tolerance-Based Fairness Auditing: Violation Certification and Sensitivity Screening
Machine Learning
Summary
When AI systems treat different groups unfairly, people want to check if this unfairness is acceptable or too much. The authors present a new way to test if an AI’s bias is above a certain allowed level depending on the rules or situation. They created two types of tests: one that safely confirms if there is unfairness without false alarms, and another that catches unfairness early, even if some false alarms happen. They showed how these tests work by analyzing a real-world AI tool used in criminal justice.
What this means in practice
- •For machine learning audit teams: Test if AI system bias violates legally or ethically allowed unfairness limits using a rigorous statistical framework.
- •For regulatory compliance teams: Screen AI tools early for unacceptable fairness violations to meet regulatory requirements and prevent ethical risks.
Tested on one dataset.
Authors
Jie Tang, Chuanlong Xie, Lixing Zhu
Abstract
As artificial intelligence is increasingly deployed, algorithmic unfairness has raised growing concerns and intensified demands for transparent fairness auditing. In practice, the tolerable degree of algorithmic unfairness depends on the specific legal, ethical, or application context. Given a prespecified tolerance threshold, an important statistical question is how to determine whether a group disparity exceeds the allowable tolerance across different auditing objectives. To address this problem, we develop a unified tolerance-based fairness auditing framework for two complementary auditing objectives: violation certification, which prioritizes control of false violation declarations, and sensitivity screening, which prioritizes reducing missed violations. For the first objective, we develop a constrained empirical likelihood test for formal settings that uses least-favorable-point calibration and can be combined with false flagging rate control for simultaneous subgroup auditing. For the second objective, we develop split empirical likelihood and adjusted split empirical likelihood tests using an adaptive boundary-proxy principle for early-warning settings. Numerical experiments show the distinct error-control--sensitivity trade-offs of these procedures. A COMPAS analysis illustrates the framework in predictive fairness auditing.