Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification

Joy Buolamwini and Timnit Gebru evaluate three commercial gender classifiers on the 1,270-person Pilot Parliaments Benchmark. Table 4 reports errors of 20.8–34.7% for darker-skinned female subjects and 0–0.8% for lighter-skinned male subjects in these tests.^[1]

Why the aggregate can mislead

Consider a hypothetical test with 900 cases in group A and 100 in group B. A system gets 95% of A and 50% of B correct: 855 + 50 = 905 correct, or 90.5% overall. The headline score conceals the difference between groups. This original arithmetic example illustrates aggregation; it is not another result from the paper.

Task and label boundaries

The study tests gender classification, not identity matching. Its binary perceived-gender labels do not capture gender’s full complexity (printed p.6). These findings cannot supply a police identification error rate.^[1]

Compare NIST’s identity-recognition evaluation and the Williams incident. Social Alignment asks how measurement relates to decisions about permissible use.

Historical context

The proceedings record specifies February 23–24, 2018; chronology uses the conference month. Selected original methods, labels and Table 4 checked, including rendered p.9. Code, dataset reproduction, supplements and later vendor performance were not audited.^[1]

Follow-up evidence

The 2019 follow-up revisits the benchmark and disclosure process. Inioluwa Deborah Raji supplies a reading entry into that work and an original interview about auditing. Keep the original study and later measurements distinct.