Research Article / Proceedings of the 1st Conference on Fairness, Accountability and Transparency
Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification
Abstract This study presents an approach to evaluate bias in automated facial analysis algorithms and datasets with respect to phenotypic subgroups. Using the dermatologist-approved Fitzpatrick Skin Type classification system, the researchers found severe gender and skin-tone bias in commercial gender classification systems, with darker-skinned females showing error rates of up to 34.7% compared to just 0.8% for lighter-skinned males. This finding reveals intersectional discrimination in AI systems and demands urgent attention for building fair, transparent, and accountable facial analysis algorithms.
Joy Buolamwini, Timnit Gebru / Read paper