PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Luhang Sun, Mian Wei, Yibing Sun, Yoo Ji Suh, Liwei Shen, Sijia Yang

arXiv:2305.10566v1cs.CVcs.AIcs.CY

TL;DR

Generative AI may reproduce or amplify gender bias in visual media, while DALL·E 2's algorithms and training materials remain opaque. This study audits representational and presentational biases using occupational prompts and comparisons with census and Google Images benchmarks, finding evidence of bias amplification in DALL·E 2.

  • Problem

    The study addresses concerns that generative AI may reproduce or amplify gender bias in visual media, despite limited transparency about DALL·E 2's algorithms and training materials.

    Source: Introduction · p. 2

  • Method

    The authors audit representational and presentational biases in DALL·E 2 images using occupation-based prompts and comparisons with census data and Google Images.

    Source: Introduction · pp. 2–3

  • Results

    DALL·E 2 images showed both representational and presentational gender biases, with evidence of bias amplification relative to benchmark data.

    Source: Introduction · p. 3

  • Takeaways & Limitations

    The authors call for feminist interventions to prevent bias-laden AI-generated visuals from permeating the media ecology.

    Source: Introduction · p. 3

    DALL·E 2's algorithms and training materials remain a “black box.”

    Source: Introduction · p. 2

Abstract

from arXiv · show

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.

Introduction

This study audits representational and presentational gender biases in DALL·E 2 images, addressing concerns that generative AI can reproduce or amplify gender bias in visual media. It compares generated images with labor statistics and Google Images, finding evidence of bias amplification in DALL·E 2.

Introduction

  • Because DALL·E 2's algorithms and training materials are a “black box,” systematic auditing is needed to assess its gender biases.The authors call for feminist interventions to prevent bias-laden AI-generated visuals from permeating the media ecology.Source: Introduction · pp. 2–3
  • Generative AI can directly create media content, raising concern that gender-biased images could enter and affect the media ecology.The authors call for auditing before DALL·E 2's widespread adoption systematically skews visual media along gender lines.Source: Introduction · p. 2
  • The study examines both unequal representation of women and men and differences in how they are presented, an area that has received less attention in algorithmic communication research.The authors document the prevalence and magnitude of both bias types in DALL·E 2 images.Source: Introduction · p. 2
  • The authors benchmark occupational gender bias using 2021 census labor statistics and Google Images.They derive prompts from a list of occupations and computationally extract visual features from DALL·E 2 images.Source: Introduction · p. 3
  • DALL·E 2 showed evidence of gender-bias amplification relative to benchmark data, while Google Images exhibited signs of bias countering.The study reports both representational and presentational biases in AI-generated images.Source: Introduction · p. 3

Algorithmic biases, generative AI, and gender bias amplification

Prior research has documented algorithmic gender bias in areas such as occupational roles and image search, while audits of generative AI have tended to focus on text. Because generative AI creates content directly, biased images may feed back into the media ecology.

Algorithmic biases, generative AI, and gender bias amplification

  • Algorithmic systems can encode social biases, and their increasing use risks threatening progress in global equality and human rights.The cited discussion frames algorithmic power as unevenly distributed and capable of amplifying existing disparities.Source: Algorithmic biases, generative AI, and… · pp. 3–4
  • Prior research finds that occupational image-search results can underrepresent women and exaggerate gender stereotypes, including across 57% of 105 jobs.Searches for “CEO” have also returned substantially more images of men than women.Source: Algorithmic biases, generative AI, and… · p. 4
  • Generative AI directly participates in content production, and its outputs may become training data for future or other systems, creating a possible feedback loop that reproduces gender bias.The passage notes this risk if no screening is applied.Source: Algorithmic biases, generative AI, and… · p. 5
  • Existing audits of generative AI tend to examine textual output, and image-generative systems such as DALL·E 2 remain relatively understudied.The authors argue that timely examination is warranted as generative AI products update rapidly.Source: Algorithmic biases, generative AI, and… · p. 5

Representational gender bias in mass media and its negative effects

This section reviews unequal and stereotypical portrayals of women in mass media and their potential harms. It defines representational bias and sets out comparisons of occupational portrayals against workforce statistics and Google Images.

Representational gender bias in mass media and its negative effects

  • The literature review describes women as more likely to be shown in domestic, dependent, or service-oriented roles and less likely to appear in professional or science jobs.These patterns include depictions of women as nonprofessionals, homemakers, and sexual gatekeepers.Source: Representational gender bias in mass… · p. 6
  • Unequal visibility in media may reinforce status beliefs that women are less competent and important than men.The section also describes prior findings that stereotypical portrayals can affect women's performance and career choices.Source: Representational gender bias in mass… · pp. 5–6
  • The authors define representational bias as unequal representation of men and women across media settings, especially overrepresentation of women in stereotypically feminine roles.They describe women's underrepresentation and stereotypical occupational portrayals in mass media.Source: Representational gender bias in mass… · p. 6
  • Prior research on Google image searches found women underrepresented across occupations and particularly below their workforce share in more than half of audited occupations.The representation deficit was more severe in male-dominated occupations, indicating amplification of representational bias by Google Images.Source: Representational gender bias in mass… · p. 7
  • The authors expect DALL·E 2 to underrepresent women in male-dominated occupations and overrepresent them in female-dominated occupations relative to census data and Google Images.They categorize jobs as male-dominated, female-dominated, or relatively equal to assess potential amplification.Source: Representational gender bias in mass… · p. 7

Presentational gender biases: Emotions and Gestures

This section distinguishes presentational bias from unequal representation and reviews gendered portrayals of emotion, gesture, and facial orientation. The authors propose auditing smiles, calmness, and face-pitching in DALL·E 2 images across occupations and against Google Images.

Presentational gender biases: Emotions and Gestures

  • Prior media research describes women as more often shown with positive emotions and smiling, reflecting cultural expectations of positive, “lady-like” behavior.Women politicians are also described as more likely than men to display visible positive emotions on television.Source: Presentational gender biases: Emotions… · pp. 7–8
  • Women are often portrayed as subordinate or passive through facial expressions and gestures, while men are depicted as dominant and confident.The section connects such stereotypes with beliefs about men's competence and status worthiness.Source: Presentational gender biases: Emotions… · p. 8
  • Upward-pitched faces convey more authority and dominance than downward-pitched faces, which may contribute to gendered differences in facial orientation.The section notes that media often present men as dominant and women as subordinated.Source: Presentational gender biases: Emotions… · p. 8
  • The study examines presentational biases in smile, calmness, and face-pitching, including whether they vary by occupation and are amplified relative to Google Images.The authors describe presentational bias as receiving less attention in generative-AI research.Source: Presentational gender biases: Emotions… · pp. 8–9
  • The authors propose testing whether women are more likely to smile and pitch downward than men in DALL·E 2 images, especially in female-dominated occupations, and whether these patterns exceed those in Google Images.They also hypothesize that women will be less likely to show calmness, with amplification strongest in female-dominated occupations.Source: Presentational gender biases: Emotions… · p. 9

Datasets

The study benchmarks DALL·E 2 images against 2021 CPS occupational data and Google Images, using 100 images for each of 153 occupations in both image datasets.

Datasets

  • The study evaluates prevalence, magnitude, and types of gender bias in DALL·E 2 images generated by occupation against census data on occupational gender segregation and Google Images, benchmarks for current disparities and common online images.Source: Datasets · p. 9
  • The CPS data were filtered and screened to produce a final set of 153 occupations with sufficient detectable-face images.Occupations without gender-disparity information were removed; the study then selected the top half by employee count in each category and excluded occupations yielding too few detectable faces.Source: Datasets · pp. 9–10
  • Google Images were collected through occupation-specific searches, and DALL·E 2 images were generated for the same occupations until each had 100 detectable-face images.Each image dataset contains 15,300 images.Source: Datasets · p. 10

Measures

The study extracts face and visual-feature measures from images, validates automated annotations against human coding, and models representational and presentational gender patterns across occupations and sources.

Measures

  • Machine-coded face, gender, smile, and emotion annotations were validated against annotations from three undergraduate coders blind to study hypotheses, yielding average Precision .85, Recall .92, and F-score .87.Krippendorff’s alpha was .90.Source: Measures · p. 10
  • For multi-face images, the analysis selected the detected face with the largest bounding-box area as most prominent; it categorized 153 occupations in the 2021 CPS dataset by female employment share.The cited passage gives a cutoff of less than 33.3% females (N = 57).Source: Measures · pp. 10–11
  • One- and two-proportion Z-tests compare female-image shares with CPS and Google Images, while mixed models analyze smile, calmness, and pose pitch.Occupation type is included as a random intercept; models test gender, occupation type, image source, and their interactions.Source: Statistical Analysis · p. 11
  • Presentational-bias models test gender, occupation type, source, and their interactions, using logistic mixed models for smiles and linear mixed models for continuous outcomes.The continuous outcomes include calmness and pose pitch; DALL·E 2 is the reference source for the source factor.Source: Statistical Analysis · p. 11

Representational gender biases in DALL·E 2: systemic underrepresentation and stereotypical

DALL·E 2 underrepresents women in male-dominated and relatively-equal occupations while overrepresenting them in female-dominated occupations; Google Images appears to counteract this pattern.

Representational gender biases in DALL·E 2: systemic underrepresentation and stereotypical

  • 38.4% of DALL·E 2 images were female and 61.6% male, compared with 46.4% female and 53.6% male among Google Images.Each source contributed 15,300 images.Source: overrepresentation · p. 12
  • Women were significantly underrepresented in DALL·E 2 images for most male-dominated and relatively-equal occupations, but significantly overrepresented in female-dominated occupations.The authors note that the latter pattern may reinforce occupational gender segregation.Source: overrepresentation · p. 12
  • Google Images tended toward higher female shares in male-dominated and lower female shares in female-dominated occupations, counteracting the DALL·E 2 pattern.The DALL·E 2–Google comparison also found women overrepresented in female-dominated occupations and underrepresented in male-dominated and relatively-equal occupations in DALL·E 2.Source: overrepresentation · p. 12

Smile

This section examines occupational gender representation and presentation in DALL·E 2 images, including smiling, calmness, and head pitch, with comparisons to Google Images. It reports the relevant model-result tables, sample sizes, and modeling conventions.

Smile

  • Across occupations, women were more likely than men to smile in DALL·E 2 images (OR = 2.19, 95% CI = [2.01, 2.39], p < .001), with a larger gender disparity than in Google Images.The source comparison was OR = 0.81, 95% CI = [0.73, 0.91], W(1) = 13.84, p = .001.Source: Smile · p. 13
  • Smiling faces were least likely in male-dominated jobs; smiling bias—higher proportions of smiling women than men in female-dominated and relatively-equal occupations versus male-dominated jobs—was more pronounced in DALL·E 2 images than Google Images.Source: Smile · p. 13

Calmness

  • Across occupation categories in DALL·E 2 images, men were more likely than women to display calmness (b = –11.94, p < .001), and the disparity was more severe than in Google images (b = 3.45, F(1, 30567.4) = 15.18, p = .001).Source: Calmness · p. 13

Pose pitch

  • Within DALL·E 2 images, women showed a stronger tendency to pitch downward than men, particularly in female-dominated versus male-dominated jobs (Gender × Female Dominated: b = –3.60, p = .023).Source: Pose pitch · p. 14

Discussion

  • DALL·E 2 systematically underrepresented women in male-dominated jobs and overrepresented them in images portraying female-dominated occupations, exhibiting more severe representational bias than Google Image, which appeared to counter existing occupational gender segregation.Source: Discussion · pp. 14–15
  • Because model development remains opaque and the audit is descriptive, the authors call for ongoing collaborative monitoring and further research on how exposure to AI-generated images affects perceptions.They also recommend collaboration on data sharing, model auditing, and debiasing.Source: Discussion · pp. 15–16
  • Across occupations, DALL·E 2 produced more images of smiling women and calm men than Google Images, and more often showed women smiling and pitching their heads down in female-dominated jobs.The authors describe these patterns as risks of amplifying presentational biases in a gender-stereotypical direction.Source: Discussion · pp. 15–16
  • The authors urge generative-AI developers and stakeholders to take responsibility and incorporate feminist epistemology into practice, because visibility matters.They frame this as an effort to reduce gender bias.Source: Discussion · p. 16

Limitations

  • The analysis focused on the largest face in multi-face images and used a 50% detection threshold, potentially losing image-level nuance.The gender variable was binary because of limitations in Rekognition and census data.Source: Limitations · p. 17
  • Rekognition and human coding relied on stereotypical conceptions of gender presentation, so the study may have reinforced binary gender stereotypes.The authors recommend expanding gender categories to improve inclusivity.Source: Limitations · p. 17, Conclusion · p. 17
  • The study did not experimentally test whether exposure to gendered images affects occupational norms, beliefs, attitudes, or behavior.The authors encourage future research to address these potential effects.Source: Limitations · p. 17
  • Future work should continue auditing generative-AI gender bias and test technical, legal, administrative, and individual resistance or de-biasing strategies.The conclusion also calls for research to explore effective ways to mitigate bias.Source: Limitations · p. 17, Conclusion · p. 17

Conclusion

  • DALL·E 2 underrepresents women in male-dominated occupations and overrepresents them in female-dominated ones, while depicting women as smiling and head-down more often, especially in female-dominated occupations; its biases exceed Google Images’.The authors call for continued monitoring and research on expanding gender categories, exposure effects, and bias mitigation.Source: Conclusion · p. 17

Probability of Smile

Pose Pitch

  • The pose-pitch figure presents results from Table 4 Models 1 and 2 with 95% confidence intervals.Source: Pose Pitch · p. 30
  • Table A2 presents linear mixed model results for calmness.Source: Table A2 · p. 32
  • Models included 30,600 observations across 153 occupations in Model 1 and 27,600 observations across 138 occupations in Model 2.Model 2 excluded occupations with no female or male images in either source category; models included by-occupation random intercepts.Source: Pose Pitch · p. 32
  • Table A3 presents linear mixed model results for pose pitch.Source: Table A3 · p. 33
Loading 2305.10566v1…