What Must a Fairness Audit Report When Demographic Data Is Incomplete?

Abstract

Fairness audits are a key component of responsible machine-learning deployment. Yet what such an audit must disclose, when the protected labels it depends on are incomplete, remains unsettled. In this work, we focused on the rates a fairness audit publishes and on what an oversight reader needs beside them. We paired every published rate with a matched baseline drawn from the same audit, one hiding protected labels and one varying only the run seed. Across ACS/Folktables tasks, missingness settings that kept some protected labels moved the selected mitigation less than an ordinary rerun did. At zero protected-label access, candidates collapsed to empirical risk minimization, so the apparent exception there reflected the candidate set's composition. Equalized-odds threshold optimization most often regressed an intersectional subgroup, but that rate fell back to its baseline once we kept only the configurations an auditor would accept. Any accuracy it lost fell on the population as heavily as on the worst-off cell, so the mechanism is levelling down. The one effect that survived was a change in which cell is worst-off. Overall, our results highlight that a published audit rate should be reported with the baseline needed to interpret it, the candidate set it came from, and its intersectional effects, before it is treated as evidence about a deployed model.