Social Biases through the Text-to-Image Generation Lens

Motivation

• Generated occupation images can worsen representational fairness relative to U.S. labor statistics and Google Image Search, with near-zero women CEOs and programmers in DALLE-v2 and near-all women nurses and housekeepers in Stable Diffusion.Figure 1 illustrates the comparison for five occupations.Source: Figure 1 · p. 1, §1 · p. 1
• Because T2I models learn from massive web-crawled datasets, biases and underrepresentation in web data may reappear in generated content and reinforce existing harms [4].The paper notes that proprietary models' training datasets are not publicly available to researchers, while Stable Diffusion's data derives from open-source LAION initiatives [25] [26].Source: §1 · p. 1

Study scope and method

• The study compares DALLE-v2 and Stable Diffusion across gender, race, age, and location using occupation, personality, person, and everyday-situation prompts, with automated and crowdsourced evaluation.Occupation and personality prompts are tested both neutrally and with specified gender, age, or race; everyday situations use neutral and location-specific prompts, whose generation distance measures country representation in defaults.Source: §1 · p. 2

Results

• Both models reinforce occupational and personality stereotypes, but DALLE-v2 skews toward white, younger men and more extreme group omissions, while Stable Diffusion favors white women and is more age-balanced.Prompt expansion often diversifies results, though some occupations show gender-linked image-quality discrepancies and some cases do not improve.Source: §1 · p. 2
• For everyday situations, Nigeria, Ethiopia, Papua New Guinea, and Colombia are farthest from default generations, while the United States, Australia, and Germany are closest.India is among the farthest for Stable Diffusion only.Source: §1 · p. 2

Social Bias in Image Search

• Image-search studies found modest gender underrepresentation (50% women in U.S. BLS data corresponded to about 45% in results), with women and people of color still underrepresented in 2020.Other studies found similar biases across search engines and dimensions including geography; this work reports T2I generation biases more severe than image search.Source: §2 · p. 2

Text-to-Image Generative Models

• DALLE-v2 uses CLIP embeddings and pixel-space denoising, whereas Stable Diffusion denoises in latent space to enable high-quality generation on low-cost GPUs.The study treats them as representatives of pixel-space and latent-space generation approaches.Source: §2 · p. 2

Social bias in Text-to-Image Generative Models

• Prior T2I studies found Stable Diffusion skewed gender and skin-tone outputs, amplified intersecting stereotypes [3], while DALLE-v2 showed more similar gender presentation than CogView2 [6] and Stable Diffusion.Bianchi et al. [3] also reported that group-specific prompts, model guardrails, and expanded prompts did not lessen complex negative associations.Source: §2 · p. 4
• The study compares DALLE-v2 and Stable Diffusion v1 across people, occupations, traits, and everyday life for gender, race, age, and geography using human and automated evaluations.It also quantifies prompt-crafting effects for occupational queries, beyond prior example-based evidence.Source: §2 · p. 4

3.1 Social Bias Dimensions

• The study uses binary perceived gender, seven race categories, and four age groups to analyze attributes in generated images.The authors note that these discrete categories simplify socially constructed or continuous identities and do not capture non-binary definitions of gender.Source: §3.1 · pp. 4–5

3.2 Social Bias Subjects

• Prompts cover people, occupations, personality traits, and everyday situations, with occupation comparisons referenced to 2022 U.S. BLS estimates.The study also tests occupation prompts specifying gender, race, or age; trait adjectives come from Abele et al., and location-specific situations use country names rather than country-based adjectives because the latter produced flag-dominated images.Source: §3.2 · pp. 5–6

3.3 Model and Data

• The study generated images with DALLE-v2 and Stable Diffusion V1, retaining detected-human images for analysis.It used the first 50 qualifying images per person, occupation, and trait prompt, and 250 for everyday situations and explicitly specified occupation prompts.Source: §3.3 · p. 6

3.4.1 Human Evaluation.

• Three Mechanical Turk workers labeled each image for visible humans, perceived gender, race, and age, with agreement by at least two workers required for ground truth.Images with three different responses were marked unclear and excluded; workers received FairFace [16] examples with ground-truth labels.Source: §3.4.1 · p. 6

3.4.2 Automated Evaluation.

• Automated evaluation detects humans and classifies face attributes, while comparing average CLIP [21] embeddings to assess country representations in everyday-situation images.The country comparison uses distances between default and location-specific prompt embeddings; lower distance indicates a closer representation to the default.Source: §3.4.2 · p. 6

4.1 What does a person look like in T2I generation?

• DALLE-v2 depicted more men than women (70% vs. 30%), while Stable Diffusion depicted more women than men (66% vs. 28%).The analysis covered 50 images generated with the prompt “person” and used both human and automated evaluations.Source: §4.1 · p. 6
• At least 70% of images from both models depicted white individuals, and each model omitted multiple racial groups.DALLE-v2 did not represent East Asian, Southeast Asian, or Middle Eastern individuals; Stable Diffusion did not portray Latino or Middle Eastern individuals.Source: §4.1 · p. 6
• Stable Diffusion showed more varied ages, while 76% of DALLE-v2 images depicted adults aged 18–40.Source: §4.1 · p. 6

4.2.1 Neutral Occupations.

• Only eight of 43 occupations for DALLE-v2 and seven of 43 for Stable Diffusion matched labor-statistics women’s representation within ±5%.Both models reinforced under- and over-representation in specific occupations; their women-representation correlations with labor statistics were 0.84 for DALLE-v2 and 0.87 for Stable Diffusion.Source: §4.2.1 · pp. 7–8, Figure 8 · p. 7, Figure 9 · p. 8
• Black individuals were absent from 72% of DALLE-v2 occupations and 37% of Stable Diffusion occupations, while white representation was under-represented by more than 50% in some occupations.For DALLE-v2, only Pilot and Primary School Teacher matched BLS white representation; for Stable Diffusion, only Nurse matched. Stable Diffusion also over-represented black childcare workers by 48% and under-represented black telemarketers by 21%.Source: §4.2.1 · p. 8
• Age distributions varied by occupation and model, with some occupations dominated by younger or middle-aged people and others prominently represented by people over 60.DALLE-v2’s 18–40 group reached at least 96% in five named occupations, while its 40–60 group reached at least 78% for truck driver and CEO; Stable Diffusion’s corresponding minima were 98% and 60%.Source: §4.2.1 · p. 8

4.2.2 Expanded prompts.

• Expanded gender, race, and age prompts did not reliably control depicted attributes, and Stable Diffusion ignored these specifications more often than DALLE-v2.DALLE-v2 still produced female images for 5% of male-occupation prompts and male images for at least 5% of female-occupation prompts; Stable Diffusion examples included 60% female images for “a male administrative assistant” and images of people over 60 for junior-occupation prompts at least 10% of the time.Source: §4.2.2 · pp. 8–9
• Prompt expansion may introduce new biases or image-quality drops, so the authors recommend complementary strategies such as duplicate prompting and output filters.The paper cautions that attribute prompts alone may not mitigate bias even when expanded prompts are effective.Source: §4.2.2 · p. 9

4.2.3 Image Quality Evaluation.

• Gender-skewed representations generally had greater similarity to real-world images, while DALLE-v2’s “female announcer” images showed limited diversity and were predominantly East Asian.The paper identifies lack of output diversity as a possible driver of worse image-quality scores and notes discrepancies can persist for gendered prompts.Source: §4.2.3 · p. 9

4.3 Representational bias for personality traits

• Because Stable Diffusion generated non-human images for over 50% of some trait prompts, the study focused exclusively on DALLE-v2 and used human evaluation.Source: §4.3 · p. 9
• Men are primarily associated with competence traits, while women are most strongly associated with warm traits.Examples include “intelligent,” “strong-minded,” and “rational” for men, and “affectionate,” “warm,” and “sensitive” for women.Source: §4.3 · p. 9, Table 3 · p. 9
• “Ambitious” and “determined” are most associated with Black people, while “vigorous” and “detached” are most associated with East Asian people.Source: §4.3 · p. 9
• White people are more commonly represented for several positive traits and strongly linked with “dominant” and “egoistic,” but less represented for other specified traits.Positive traits with greater representation include “competent,” “active,” “rational,” and “sympathetic”; representation is comparatively lower for “ambition,” “vigorous,” and “striving,” and lower for “detached” and “hardheaded.”Source: §4.3 · p. 9
• Caring and altruistic prompts are associated with people appearing over 60, rationality and tolerance with ages 40–60, and laziness, ambition, and perfectionism with ages 18–40.Source: §4.3 · p. 9

• The analysis compares six everyday-situation categories across 12 locations, using CLIP embedding distances between default and location-specific generations to assess country representation.Lower distance indicates a country representation closer to the default generation.Source: §4.4 · p. 9
• For events, Nigeria, Ethiopia, and Papua New Guinea have the lowest representation in both models, while Australia, Germany, and the United States are most represented.Source: §4.4 · p. 10
• Across situation prompts, Nigeria, Ethiopia, and Papua New Guinea are least represented by both models; Germany leads DALLE-v2, the United States leads Stable Diffusion, and DALLE-v2 better represents all studied countries overall.Source: §4.4 · p. 10

4.5 Human vs. Automated Evaluation

• The section reports correlation results comparing human and automated assessments of occupations and traits.The assessments were collected across gender, race, and age dimensions and are presented in Tables 4 and 5.Source: §4.5 · p. 10
• Correlation coefficients exceeded 0.9 for all groups except white individuals in the occupation category.The assessments covered demographic dimensions including gender, race, and age.Source: Table 4 · p. 10, Table 5 · p. 10, §4.5 · p. 10
• Meaningful correlation scores could not be computed for people younger than 18 or older than 60, or for race groups other than black and white.Because these groups were significantly under-represented in generations from both models, automated evaluation performance for them remains inconclusive.Source: §4.5 · p. 10

4.6 Limitations

• The analysis does not deeply assess non-binary gender or underrepresented communities, and whether prompt expansion helps them—and at what cost—remains uncertain.The authors note that example-based evidence suggests these groups are poorly represented in neutral prompts; expanded prompts may fail to address image-quality discrepancies or complex associations requiring qualitative evaluation.Source: §4.6 · p. 11
• Country-specific prompts can still portray some countries as economically poor, while non-binary or religious-group generations risk harmful stereotypes [3].The authors argue that these associations must be studied to assess how far prompt expansion can mitigate representational fairness concerns.Source: §4.6 · p. 11
• The evaluation covers only DALLE-v2 and Stable Diffusion v1, leaving proprietary models such as Imagen and newer models such as Stable Diffusion v2 for future study.The authors call for evaluation of models continuing to be released and deployed in real-world applications.Source: §4.6 · p. 11

5 CONCLUSION

• DALL-E V2 and Stable Diffusion v1 showed significant biases across gender, race, age, and geographical location, exacerbating patterns relative to recent BLS labor statistics.The study used human and automated evaluations of prompts covering occupations, personality traits, everyday situations, and a general “person” prompt.Source: §5 · p. 12
• Prompt expansion strategies could diversify generated content, but could also lead to variations in image quality.Source: §5 · p. 12
• Images depicting everyday situations underrepresented some countries and overrepresented others.Source: §5 · p. 12
• The study presents its results and evaluation method as resources for evaluating and building generative models with greater attention to responsible development and representational fairness.The authors plan to explore more mitigation strategies.Source: §5 · p. 12

Occupation prompts

• Table 7 lists the personality-trait prompts used in the study, following adjectives proposed in previous work [2].The passage does not provide occupation-prompt details.Source: Table 7 · p. 13

Occupational comparisons

• A separate figure shows Image Search Engine, DALLE-v2, and Stable Diffusion images generated from the prompt “Engineer.”Source: Figure 17 · p. 14
• Occupational supplements include BLS 2022 comparisons for occupation prompts, including correlation and gender- and race-proportion plots across DALLE-v2 and Stable Diffusion.Source: Table 8 · p. 14, Figure 18 · p. 14, Figure 19 · p. 14

Occupational image quality

• The appendix reports FID scores for gendered occupational prompts separately for DALLE-v2 and Stable Diffusion, with lower scores indicating closer image distributions to Image Search images.Source: Figure 23 · p. 17, Figure 24 · p. 17

Personality traits

• Two figures show race distributions for positive and negative personality traits.Source: Figure 25 · p. 18, Figure 26 · p. 18

Geographical representations

• Heat maps show DALLE-v2 and Stable Diffusion representations for places, food, and institutions.Source: Figure 27 · p. 19, Figure 28 · p. 19, Figure 29 · p. 20
• Additional heat maps cover community and clothing categories for both models, while country charts identify the least represented countries across situation prompts.Source: Figure 33 · p. 21, Figure 34 · p. 21, Figure 35 · p. 21
• Separate charts identify the most represented countries across situation prompts for DALLE-v2 and Stable Diffusion.Source: Figure 38 · p. 22, Figure 40 · p. 22

Human evaluation

• The appendix also includes the Amazon Mechanical Turk questionnaire.Source: Figure 41 · p. 23