# Smiling Women Pitching Down: Auditing Representational and Presentational Gender Biases in Image Generative AI

**Authors:** Luhang Sun, Mian Wei, Yibing Sun, Yoo Ji Suh, Liwei Shen, Sijia Yang  
**arXiv:** [2305.10566](https://arxiv.org/abs/2305.10566)  
**Date:** 2023-05-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.10566v1/

## Key points

- Study title: "Smiling Women Pitching Down: Auditing Representational and Presentational Gender Biases in Image Generative AI"
- Examined gender biases in AI-generated images, focusing on representational and presentational biases
- Analyzed 15,300 DALL-E 2 images across 153 occupations using computational algorithms
- Identified more pronounced biases in DALL-E 2 images compared to Google Images
- Found underrepresentation of women in male-dominated professions and overrepresentation in female-dominated roles
- Noted a trend of depicting more women with smiling faces and downward-pitching heads, especially in female-dominated fields
- Emphasized the lack of systematic auditing for gender biases in generative AI models' imagery
- Compared findings against census labor statistics and Google Images to assess bias amplification
- Urged for feminist interventions to address bias-laden AI-generated imagery perpetuating stereotypes and shaping societal perceptions through media channels

## Summary

In their study titled "Smiling Women Pitching Down: Auditing Representational and Presentational Gender Biases in Image Generative AI," authors Luhang Sun, Mian Wei, Yibing Sun, Yoo Ji Suh, Liwei Shen, and Sijia Yang delve into the examination of gender biases present in AI-generated images produced by models like DALL-E 2. The researchers focused on two specific types of occupational gender biases - representational biases and presentational biases - and conducted a comprehensive analysis of 15,300 DALL-E 2 images across 153 different occupations. By employing computational algorithms for auditing purposes, the researchers were able to identify more pronounced representational and presentational biases within DALL-E 2 images compared to those sourced from Google Images. Their investigation revealed significant disparities in the representation of genders across various occupational fields. Specifically, DALL-E 2 was found to underrepresent women in male-dominated professions while overrepresenting them in female-dominated roles. Furthermore, the study highlighted a tendency for DALL-E 2 images to depict more women than men with smiling faces and downward-pitching heads, particularly in occupations traditionally dominated by females. The widespread excitement surrounding the capabilities of generative AI models to create visually stunning images based on textual prompts has led to a notable lack of systematic auditing when it comes to identifying potential gender biases within these images. To gauge the extent of bias amplification, the researchers compared their findings against 2021 census labor statistics and Google Images. The study's findings underscore the importance of implementing feminist interventions to mitigate bias-laden AI-generated imagery from perpetuating stereotypes and influencing societal perceptions through media channels. This research sheds light on the critical need for ongoing scrutiny and evaluation of AI-generated content to ensure fair and accurate representation across all genders in visual media.