Bias Audits Detect Bias but Disagree on Ranking: Evidence from Ten Instruments and Ten Frontier Models
Source
arxiv.org
Author
William Guey, Pierrick Bougault, Wei Zhang, Vitor D. de Moura, Jos\'e O. Gomes
Date
Context
A bias test can detect a problem without establishing which model is least biased overall. In a comparison of ten audit tools on ten models, eight tools detected bias, yet their model rankings agreed no better than chance. Adding weaker models improved reliability within individual tools without restoring agreement between tools. The authors interpret this as evidence that the audits measure different aspects of bias. A ranking therefore needs to be read alongside what the audit asks models to do; a strong result on one format cannot stand in for a general fairness ranking.
Key takeaways · AI-distilled
Running ten bias-audit tools on the same ten frontier models, eight of ten detected bias with confidence intervals clear of zero, but two widely cited direct-probe benchmarks were saturated because frontier models now answer neutrally.
In this study, the audit tools disagreed on model rankings about as much as chance would predict. Detecting a bias within one test therefore did not produce a consistent answer to which model was least biased.
Adding six deliberately weaker models to the panel restored each tool's internal reliability but still did not fix cross-tool ranking agreement, suggesting the audits measure different constructs rather than one construct with noise.
The direction of detected bias depended on audit format: forced-choice decision tools mostly over-corrected toward women and toward working-class candidates in 273 of 278 hiring decisions, while free-generation and default coreference tests stayed stereotype-congruent.
The authors conclude that an audit can detect bias within the behavior it tests, but its score alone does not justify ranking one model as less biased overall. The choice of test changes what kind of bias is measured.
articleEvalDetectBench: A Benchmark for Measuring Evaluation Awareness in Frontier Language ModelsXinning Li, Kemunto Ochwang'i, Aryasomayajula Ram Bharadwaj, Alexandra Souly, Robert Kirk
articleWho Do Language Models Think Is Competent? A Mechanistic Analysis of Occupational BiasKeren Fuentes, Aaron Mueller
articleThe Double Measurement Confound in Agent Benchmarks: De-Scaffolding, Ground-Truth Scoring, and Reliability Beyond the MeanYonghong Zhang, Shadi Motaali, Vu Phong Dinh, Avin Piroutiniya, Jorge E. L\'opez de Vergara, Luis de Pedro, Ricardo Correia, Isabel M. Parra, Yong Xie