Model-agnostic auditing: a lost cause?

Tools for interpretable machine learning (IML) or explainable artificial intelligence (xAI) can be used to audit algorithms for fairness or other desiderata. In a black-box setting without access to the algorithm’s internal structure an auditor may be limited to methods that are model-agnostic. These methods have severe limitations with important consequences for outcomes such as fairness. Among model-agnostic IML methods, visualizations such as the partial dependence plot (PDP) or individual conditional expectation (ICE) plots are popular and useful for displaying qualitative relationships. Although we focus on fairness auditing with PDP/ICE plots, the consequences we highlight generalize to other auditing or IML/xAI applications. This paper questions the validity of auditing in high-stakes settings with contested values or conflicting interests if the audit methods are model-agnostic.