Signal The Audit Decides the Verdict: Instrument Effects Rival Demographic Bias in LLM Decision Audits

Summary

This study shows that how an audit of large language model decisions is designed can matter as much as demographic bias itself. The researchers tested five models using 40,726 hiring, lending and medical-triage requests that varied only in applicant name. None of the 36 planned bias contrasts survived statistical correction. The models proved highly sensitive to whether an audit was transparent and to the order in which candidates were presented. That sensitivity was sometimes as large as differences tied to demographic attributes. This suggests that audit verdicts reflect audit construction more than they reflect underlying bias.