[論文レビュー] Smiling Women Pitching Down: Auditing Representational and Presentational Gender Biases in Image Generative AI

Luhang Sun, Mian Wei+4 more|arXiv (Cornell University)|May 17, 2023

Ethics and Social Impacts of AI被引用数 14

ひとこと要約

この論文は、153の職業をカバーする15,300のプロンプトにわたり、DALL-E 2画像のジェンダー・バイアスを監査し、国勢調査データとGoogle Imagesと比較して表象的および表現的バイアスを評価します。男性支配の分野で女性の過小評価、女性支配の分野で過大評価、さらに笑顔と頭の角度の偏りを発見します。

ABSTRACT

Generative AI models like DALL-E 2 can interpret textual prompts and generate high-quality images exhibiting human creativity. Though public enthusiasm is booming, systematic auditing of potential gender biases in AI-generated images remains scarce. We addressed this gap by examining the prevalence of two occupational gender biases (representational and presentational biases) in 15,300 DALL-E 2 images spanning 153 occupations, and assessed potential bias amplification by benchmarking against 2021 census labor statistics and Google Images. Our findings reveal that DALL-E 2 underrepresents women in male-dominated fields while overrepresenting them in female-dominated occupations. Additionally, DALL-E 2 images tend to depict more women than men with smiling faces and downward-pitching heads, particularly in female-dominated (vs. male-dominated) occupations. Our computational algorithm auditing study demonstrates more pronounced representational and presentational biases in DALL-E 2 compared to Google Images and calls for feminist interventions to prevent such bias-laden AI-generated images to feedback into the media ecology.

研究の動機と目的

• Representational gender bias (occupational distribution of women vs. men) in DALL-E 2 images across 153 occupations.
• Presentational gender bias (facial expressions and head pose) in DALL-E 2 images across occupations.
• Benchmark DALL-E 2 biases against 2021 census labor statistics and Google Images.
• Provide evidence and call for feminist interventions to mitigate bias in AI-generated imagery.

提案手法

• Compile a dataset of 15,300 DALL-E 2 images generated from prompts spanning 153 occupations.
• Quantify representational bias by comparing gender representation to census labor statistics.
• Quantify presentational bias by analyzing facial expressions (smiling) and head pitch (downward) by gender and occupation.
• Benchmark biases against Google Images to assess amplification or attenuation of biases.
• Develop and apply computational auditing algorithms to quantify bias characteristics in generated images.
• Discuss implications for media ecology and potential feminist interventions.

実験結果

リサーチクエスチョン

• RQ1 Do DALL-E 2 generated images underrepresent women in male-dominated occupations and overrepresent them in female-dominated ones?
• RQ2 Do DALL-E 2 images depict more women than men with smiling faces and downward-pitching heads, particularly in female-dominated occupations?
• RQ3 How do DALL-E 2 biases compare to Google Images in terms of representational and presentational biases?
• RQ4 What interventions could mitigate bias-laden AI-generated imagery influencing media ecosystems?

主な発見

• DALL-E 2 underrepresents women in male-dominated fields and overrepresents them in female-dominated occupations.
• DALL-E 2 images tend to depict more women than men with smiling faces and downward-pitching heads.
• Biases are more pronounced in DALL-E 2 than in Google Images in both representational and presentational dimensions.
• Bias amplification suggests potential feedback into media ecology requiring feminist interventions.