Interpretable Interfaces for Algorithmic Auditing

Abstract

As machine learning algorithms are used in increasingly consequential settings, the idea of auditing algorithms, or evaluating them in a structured manner against specific fairness or safety criteria, is gradually gaining traction. With a lack of federal regulation in the United States and industry standards currently emerging, advocates for more widespread, comprehensive auditing must navigate a complex technical and policy ecosystem. This thesis investigates the current landscape of algorithmic auditing in the United States, examines the potential for applications of interpretable machine learning to create more comprehensive audits, and develops ExplanAudit, a software tool to incorporate interpretability while supporting scalability, accessibility, and visibility for this work. This is accomplished through a user study of n=15 algorithmic auditors based in the United States with both qualitative and demonstration components. We find that the population of auditors appears to avoid a known pitfall of human-explanation interaction, in that they likely do not become overconfident as a result of exposure to the explanations; as a result, our toolkit supports a variety of interpretable methods for auditing and is available online. Future work could expand this study to a larger sample of auditors, extend the software to train new auditors, or apply the tool to audit specific behaviors of open-source models.