Diagnosing Gender Bias in Image Recognition Systems

Area of Science:

• Computer Vision
• Artificial Intelligence Ethics
• Gender Studies

Background:

• Image recognition systems promise scalable image analysis without expert input.
• Machine learning models are known to produce biased outputs, necessitating investigation.
• Gender bias in AI is a critical concern impacting fairness and representation.

Purpose of the Study:

• To evaluate gender biases in commercial image recognition platforms.
• To analyze bias using images of U.S. politicians from Congress and Twitter.
• To understand the implications of encoded biases on women's visibility and stereotypes.

Main Methods:

• Utilized photographs of U.S. members of Congress and their Twitter images.
• Employed crowdsourced validation to assess image recognition system outputs.
• Analyzed the types and frequency of labels generated for male and female politicians.

Main Results:

• Commercial systems generated correct yet biased labels, selectively reporting subsets of true labels.
• Images of women received three times more annotations related to physical appearance compared to men.
• Women in images were recognized at substantially lower rates than men.

Conclusions:

• Commercial image recognition systems exhibit significant gender bias.
• Encoded biases negatively affect women's visibility and reinforce harmful gender stereotypes.
• Biased AI systems limit the validity of insights derived from image data.