Employers' recent turn to AI to screen job applications raises concerns about algorithmic bias. This concern, in turn, has fueled efforts to "blind" algorithms to various applicant characteristics, such as, e.g., gender. Whether efforts to "blind" algorithms actually work endures as an empirical question.
Data on this question are analyzed in a recent paper, Blinding Gender: An Algorithmic Bias Experiment. In it, Charlotte Alexander (Georgia Tech--Business) et al. lever 9,000 recommendation letters submitted by applicants to medical residency programs in the service of testing "blinding" efforts.
The paper's core finding emphasizes that "even after blinding, our best performing model successfully predicts applicant gender 63% of the time, indicating that recommendation letters retain implicit linguistic patterns that continue to reveal identity." The authors conclude that "blinding strategies fail as a mechanical matter, in that they do not actually blind AI models to applicants’ gender" as the recommendation letters retain linguistic patterns that reveal applicants' identities (or here, gender). To be sure, as a naive ex ante prediction baseline should yield approximately 50%, whether the 63% rate found in the paper is "enough" to safely scaffold a conclusion that "blinding" efforts "fail" remains debatable. Regardless, and any quibbles notwithstanding, an excerpted abstract follows.