Skip to content
Longterm Wiki
Back

Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification

web
Semantic Scholar·semanticscholar.org/paper/Gender-Shades:-Intersectional-A...

Credibility Rating

4/5
High(4)

High quality. Established institution or organization with editorial oversight and accountability.

Rating inherited from publication venue: Semantic Scholar

Landmark study by Buolamwini and Gebru demonstrating that commercial facial recognition systems exhibit significant accuracy disparities across gender and skin tone, foundational for AI fairness and bias research.

Metadata

Importance: 85/100conference paperprimary source

Summary

This paper by Joy Buolamwini and Timnit Gebru evaluates commercial facial analysis systems and finds substantial accuracy disparities based on gender and skin tone intersections. Darker-skinned females were misclassified at rates up to 34% higher than lighter-skinned males. The study introduced the Fitzpatrick skin tone scale as an evaluation framework and spurred major AI companies to improve their systems.

Key Points

  • •Commercial gender classification systems showed error rates up to 34.7% higher for darker-skinned women compared to lighter-skinned men.
  • •Introduced intersectional analysis combining gender and skin tone (Fitzpatrick scale) to evaluate AI system fairness.
  • •Evaluated three major commercial facial analysis APIs from Microsoft, IBM, and Face++.
  • •Demonstrated that benchmark datasets used to train these systems were not representative of diverse populations.
  • •Became a foundational paper in algorithmic fairness, prompting industry responses and policy discussions around AI bias.

1 FactBase fact citing this source

EntityPropertyValueAs Of
Algorithmic Justice LeaguepublicationGender Shades study (2018) by Buolamwini and Gebru demonstrated intersectional accuracy disparities in commercial facial recognition from IBM, Microsoft, and Face++. Dark-skinned women had highest error rates; light-skinned men had lowest. Over 4,900 citations on Semantic Scholar.Mar 2026

Cached Content Preview

HTTP 202Fetched Apr 7, 20260 KB
JavaScript is disabled

 In order to continue, we need to verify that you're not a robot.
 This requires JavaScript. Enable JavaScript and then reload the page.
Resource ID: kb-800d31ec35899088