New Research Quantifies Limits of AI Fairness Auditing

New academic work explores the inherent challenges in auditing AI systems for fairness, particularly when dealing with black-box models. The study frames fairness auditing as a strategic game between a company, which has unlimited computational power to optimize its model, and a budget-constrained auditor. This setup reveals that even with dedicated audit resources, there are fundamental limits to how effectively manipulation can be prevented. The researchers quantified the minimum level of post-audit manipulation that remains unavoidable, even when auditors have a fixed budget or require a certain level of fairness approximation. They derived mathematical lower bounds for demographic parity deviation, which illustrate that while increased audit resources can reduce manipulation, they cannot eliminate it entirely. Empirical tests using simple audit-set construction heuristics on linear and neural network classifiers support these theoretical findings, underscoring the persistent challenge in achieving complete fairness certification.