PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative AI Produces and Transmits Beauty Bias

AJ Alvero, Deepak Varuvel Dennison, Alexandra Crespin, Heeyon Kim, Alejandra Regla-Vargas, Avani Akerkar, Jasmine McNealy, René Kizilcec

arXiv:2610.22553v1cs.CYcs.CV

TL;DR

This paper asks whether generative AI contributes to socially patterned beauty standards and whether generated faces change evaluations of real people. Across a generator audit and hiring experiment, it finds that generated faces are judged more attractive than photographed faces and can change how real people are judged.

  • Problem

    Whether exposure to AI-generated faces changes consequential evaluations of real people remains unclear, despite evidence that highly selected or edited imagery can affect appearance judgments.

    Source: Introduction · p. 3

  • Method

    The authors audit 27 text-to-image systems using common prompts, compare generated faces with photographs using automated and human ratings, and run a hiring experiment with candidates paired with different face types.

    Source: Introduction · p. 3

  • Results

    Generative systems produce faces people find more attractive than real ones, distribute attractiveness along racial lines, and can change how real people are judged.

    Source: Discussion · p. 11

  • Takeaways & Limitations

    Generative systems can construct represented people and attach aesthetic value to socially meaningful categories beyond isolated facial properties.

    Source: Discussion · p. 11

    The primary automated measure fell just below its prespecified human-validation threshold, so claims about human-perceived attractiveness rely on direct ratings.

    Source: Discussion · p. 11

Abstract

from arXiv · show

Societies reward people judged to be attractive by standards intertwined with gender and race. In this paper, we examine if and how generative AI may contribute to these dynamics. We audited 27 text-to-image generators released between 2022 and 2026 using simple prompts that never mention attractiveness or beauty, analyzing 51,300 generated faces and 160,196 photographs of real people across primary and supplementary comparisons. We used three automated measures and collected ratings from 550 human raters on a subset of images. Nearly all generators produced faces scoring above photographed faces under automated measures, and human raters judged the outputs of all 23 generators as significantly more attractive than photographed faces, even after correction for multiple comparisons. When focusing our analysis on gender, we find that 37% of female-prompted generated faces outscored the median professional fashion model. Across racial, ethnic, and national categories, groups associated with darker skin tones tended to receive lower attractiveness scores than groups associated with lighter skin tones. In a preregistered hiring simulation experiment (N = 570), exposure to AI-generated faces selected for high human-rated attractiveness lowered hireability and trustworthiness ratings and reduced interview advancement for an identical test set of candidates paired with photographs of real people; exposure to retouched photographs produced no such effect. Generative AI can reproduce racial hierarchies and gendered patterns of beauty and alter the context in which real people are judged.

Introduction

The paper treats beauty as socially situated evaluation rather than an intrinsic facial property, and asks how generative AI shapes represented appearances and judgments. It motivates questions about generated faces’ attractiveness, its distribution across social categories, and effects on evaluations of real people.

Introduction

  • Beauty is treated as a socially situated judgment shaped by standards that vary across evaluators and institutions yet produce patterned agreement about whose appearances are valued.The paper distinguishes these evaluative standards from an intrinsic property of a face.Source: Introduction · p. 2
  • Text-to-image systems synthesize the represented person, jointly constructing features such as age, skin tone, morphology, styling, and expression at marginal cost.This makes it feasible to populate advertisements, websites, training materials, and stock imagery with people who do not exist.Source: Introduction · p. 2
  • Attractiveness has received less direct study as an outcome of image generation, especially when prompts simply request a person or social category without specifying beauty.Prior audits documented other demographic disparities in generated representations, while this study examines appearances supplied under ordinary prompts.Source: Introduction · p. 2
  • The paper asks whether generated people are more attractive than real people, how attractiveness varies across social categories, and whether exposure changes judgments of real people.It treats computational and human ratings as measures of particular evaluative standards, not ground-truth labels of beauty.Source: Introduction · p. 3
  • The authors audit 27 text-to-image systems with prompts crossing 14 racial, ethnic, and national categories, two sex terms, and eight occupations, then test human ratings and hiring judgments.The audit compares 51,300 generated faces with 89,435 photographs using three computational measures; 550 adults rated a subset, and 570 adults took part in the hiring experiment.Source: Introduction · p. 3

Results

Generated faces scored above photographs under automated and human attractiveness assessments, while racial and ethnic categories received unequal scores. In a hiring simulation, exposure to AI-generated faces lowered evaluations of identical real candidates.

Generated faces are more attractive than photographs of real people

  • Under SCUT and MEBeauty, 24 of 27 systems scored significantly above photographs; all 23 adequately sampled generators received higher human ratings by 0.34–2.06 points.The human-rating differences remained significant after Holm correction.Source: Generated faces are more attractive… · p. 4, Figure 1 · p. 5
  • 23% of generated faces scored above the median professional fashion model, rising to 37% among female-prompted faces.Fashion models scored higher than generated faces on average under both primary automated measures; human raters also rated generated faces above photographs among both female- and male-presenting images.Source: Generated faces are more attractive… · pp. 4, 6, Figure 1 · p. 5

Racial gaps persist after measurement checks and accounting for facial characteristics

  • After adjustment, 77% of the SCUT White–Black gap and 99% of the MEBeauty gap remained; for White–Nigerian, 77% and 88% remained, respectively.CelebA/SwinFace’s fully adjusted gaps were essentially zero; these statistical adjustments are not causal decompositions, and its added attributes came from the same model as its attractiveness score.Source: Racial gaps persist after measurement… · p. 7, Table 1 · p. 7

Racialized ordering across systems and beauty standards

  • Across 27 systems, White-prompted faces scored highest on average, while Ethiopian-, Black-, and Nigerian-prompted faces occupied the bottom three across all measures; Nigerian faces ranked last.The measures disagreed about which groups ranked highest, and national and panethnic labels were not interchangeable in their scores.Source: Racialized ordering across systems and… · p. 6, Figure 2 · p. 8

Exposure to AI-generated faces lowers evaluations of real candidates

  • Directly paired with AI-generated faces, candidates received higher hireability (b = 0.162) and trustworthiness (b = 0.130) ratings, but interview advancement did not differ significantly from the real-face condition.Retouched candidates did not differ from the real-face condition on any of the three outcomes.Source: Exposure to AI-generated faces lowers… · p. 9
  • After AI-face exposure, identical real candidates received lower hireability (b = −0.146), trustworthiness (b = −0.172), and interview advancement (odds ratio = 0.768); retouched-face exposure showed no such effect.The simulation included 570 participants, and written qualifications were held fixed across conditions.Source: Exposure to AI-generated faces lowers… · pp. 7, 9, Figure 3 · p. 10

Discussion

The audit and experiment indicate that generative systems can shape both who is represented as attractive and the context in which real people are evaluated. Their effects matter beyond the images themselves, though the experiment does not identify why exposure mattered.

Discussion

  • Generative systems produce faces people find more attractive than real ones, distribute attractiveness along racial lines, and can change how real people are judged.The recurring inequalities do not require agreement on a single ideal of beauty.Source: Discussion · p. 11
  • The racial ordering persisted after accounting for constant measurement tendencies favoring White faces and for measured facial properties, but automated measures did not identify the process producing the differences.The results are more consistent with aesthetic value attached to socially meaningful categories than with an explanation based only on isolated facial properties.Source: Discussion · p. 11
  • The experiment found that people who saw highly attractive AI-generated faces later rated the same real candidates more negatively on all three outcomes.The AI faces differed from real ones in attractiveness, synthetic regularity, polish, familiarity, and demographic composition, so the experiment establishes an exposure effect but not its mechanism.Source: Discussion · p. 11
  • The experiment used one job and a deliberately high-attractiveness AI stimulus set, limiting how confidently its downstream effect generalizes to ordinary outputs from every generator.The retouched and AI-generated images were not attractiveness-matched, so the comparison does not isolate image source from treatment intensity.Source: Discussion · p. 11

Supplementary Materials

The supplementary methods describe a common generation protocol, several photographed comparison populations, and three separately reported attractiveness instruments. The model and corpus choices set the scope of those comparisons.

Image generation

  • Twenty-seven generators received the same prompts crossing 14 demographic terms, two sex terms, and eight occupations, with five images per combination.Additional conditions omitted demographic, sex, or occupation terms, bringing the corpus to 51,300 images; stylized outputs were retained.Source: Image generation · p. 16

Comparison corpora

  • The six-source pooled photographed corpus contained 89,435 usable faces, while the 10k US Adult Faces set was analyzed separately and fashion-model photographs served as an absolute calibration benchmark.The model benchmark was not treated as matched because its demographic composition, age structure, and photographic conventions differ from the generated corpus; auxiliary FFHQ and StyleGAN2 populations were not treated as reconstructions of contemporary generator training sets.Source: Comparison corpora · p. 16

Preprocessing and attractiveness measurement

  • All images were face-detected, cropped, resized, and normalized before scoring with SCUT, MEBeauty, and CelebA/SwinFace instruments.SCUT was designated the primary automated measure, with MEBeauty and CelebA/SwinFace prespecified as complements; results were reported separately because the instruments use different rating populations and operationalizations.Source: Preprocessing and attractiveness… · pp. 16–17

Human-rating validation (Study 1b)

The validation study combined human ratings of a stratified image subset with instrument-specific statistical comparisons. It also measured rendered image attributes and specified checks for demographic gaps and comparison-corpus dependence.

Human-rating validation (Study 1b)

  • Five hundred fifty U.S. adults rated 1,374 generated and photographed faces, and the 23 generators with at least 10 rated images were rated more attractive than photographs.Human-rated differences ranged from 0.34 to 2.06 points on a 1–7 scale and remained significant after Holm correction; SCUT did not meet its prespecified human-validation criterion, so human-perceived attractiveness claims rely on direct ratings.Source: Human-rating validation (Study 1b) · p. 17

Rendered attributes

  • Rendered-image attributes included estimated age, facial skin lightness, facial adiposity, blur, and 38 additional CelebA-style attribute probabilities.The SwinFace Attractive probability was excluded from the adjustment set; these measures describe rendered images, not underlying people.Source: Rendered attributes · p. 17

Statistical analysis (Study 1)

  • Study 1 compared generator and photograph distributions with Welch’s t-tests and Cohen’s d, while estimating corrected demographic contrasts and nested adjustment models.The analysis also repeated comparisons across six photographed sources, examined dispersion and upper tails, and cautioned that adjusted models are statistical adjustments rather than causal decompositions.Source: Statistical analysis (Study 1) · p. 18

Study 2 design

The preregistered study randomized U.S. adults to real, retouched, or AI-generated face exposure while they evaluated applicants for one job. Its analysis estimated direct and downstream outcomes with prespecified contrasts and sensitivity checks.

Study 2 design

  • The study recruited 570 U.S. adults and had each participant rate 20 applicant profiles, including eight condition-specific manipulation profiles and twelve common real test profiles.The AI condition used high-human-rated-attractiveness faces from Study 1b, while real and retouched conditions used the same eight identities.Source: Study 2 design · p. 18
  • All profiles described applicants for the same Customer Success Associate position, with outcomes measuring hireability, interview advancement, and trustworthiness.The three outcomes captured a job-fit judgment, a consequential selection decision, and a broader person inference.Source: Study 2 design · p. 19

Study 2 analysis

  • Mixed-effects models estimated direct and downstream effects, with a population-averaged GEE estimate used for downstream advancement after the planned model failed to converge stably.Holm-adjusted directional tests and sensitivity analyses preserved the substantive pattern; a temporal analysis tested interactions with prior manipulation exposure and trial position.Source: Study 2 analysis · p. 19

Preregistration and availability

  • The preregistration, materials, generation scripts, scoring pipelines, and analysis code are publicly available.Both human-subjects studies were reviewed and determined to be exempt.Source: Preregistration and availability · p. 19

Synthetic-image validation of the attractiveness measures

The three automated measures detect prompted attractiveness differences in synthetic images, but generated–photographed comparisons vary by instrument and benchmark. Contemporary generated faces also differ from photographed faces in distributional shape, and exceed FFHQ and StyleGAN2 under the two dedicated attractiveness measures.

Synthetic-image validation of the attractiveness measures

  • All three automated measures distinguish explicitly attractive from unattractive synthetic-image prompts in the expected direction, supporting sensitivity to manipulated attractiveness.The comparison used 26,382 usable images generated with Stable Diffusion 2.1 and 3.5 Medium.Source: Synthetic-image validation of the… · p. 19

Source-specific photographed-baseline robustness

  • SCUT and MEBeauty show positive generated–photographed effects across all six sources, while CelebA/SwinFace is source-sensitive and reverses direction for some sources.Against the 10k US Adult Faces benchmark, SCUT averaged d = 1.50 (27/27 generators positive) and MEBeauty d = 1.12 (26/27 positive); CelebA/SwinFace averaged d = −0.23 (9/27 positive).Source: Source-specific photographed-baseline… · pp. 19–20

Training-distribution and distributional-shape diagnostics

  • Contemporary generators exceed FFHQ and StyleGAN2 under SCUT and MEBeauty, while generated faces also show greater variability and disproportionate representation in the photographed distribution’s upper tail.Contemporary-versus-FFHQ effects are d = 1.25 under SCUT and d = 0.87 under MEBeauty; variance ratios are 1.53 under SCUT and 2.79 under MEBeauty.Source: Training-distribution and… · p. 20

Generator-level human-rating robustness

Image-level checks show higher human-rated attractiveness for generated than photographed faces across generators and broad racial and ethnic categories. The White–Black gap persists across prompted occupational and gender contexts, though its magnitude varies.

Generator-level human-rating robustness

  • All 23 generators’ image-mean human attractiveness ratings exceed the photographed mean, with every difference significant after Holm correction.The photographed mean was 3.10 on a 1–7 scale; the smallest corrected difference was Stable Diffusion 3 at +0.34 points (95% CI [0.04, 0.63], Holm-adjusted P = 0.026).Source: Generator-level human-rating robustness · p. 20

Human-rating demographic robustness

  • Generated faces receive higher mean attractiveness ratings than photographs in each of six shared racial and ethnic categories, with all differences remaining significant after Holm correction (all P < 0.001).These image-level checks use image means as observations and complement, rather than replace, cross-classified human-rating models.Source: Human-rating demographic robustness · p. 21

Racial gap across prompted social contexts

  • The White–Black attractiveness gap remains positive across occupational and gender prompts, although its magnitude varies by context and measure.For high-prestige occupations, the gap is +0.60 under SCUT and +0.70 under MEBeauty; under MEBeauty it is +0.86 for male-prompted versus +0.63 for female-prompted faces.Source: Racial gap across prompted social… · p. 21

Human-rating sex and attraction-target robustness

The generated–photographed human-rating difference appears across sex and attraction-target comparisons. In the hiring simulation, sensitivity analyses preserve negative AI-generated-versus-real effects, while the fielded sequence provides no evidence of a simple monotonic dose-response.

Human-rating sex and attraction-target robustness

  • Human ratings are higher for AI-generated than photographed faces among both female- and male-presenting images, including when raters assess the sex they report being attracted to.Men attracted to women rated generated male faces 0.85 points higher, and women attracted to men rated generated female faces 1.37 points higher than photographs.Source: Human-rating sex and attraction-target… · p. 21

Study 2 retouching calibration and sensitivity

  • In the n = 560 sensitivity sample, AI-generated versus real exposure remains negative for hireability, interview advancement, and trustworthiness.Estimates were b = −.154 for hireability, OR = .739 for advancement, and b = −.179 for trustworthiness; the stricter n = 495 response-quality check also retained negative effects below 0.05 after directional Holm adjustment. The retouched benchmark was calibrated to match a mean SCUT increase of +0.127, not to match the AI images’ human-rated attractiveness.Source: Study 2 retouching calibration and… · pp. 21–22

Prespecified temporal carryover analysis

  • The fielded sequence shows no evidence of a simple monotonic dose-response in downstream outcomes, and immediate-predecessor interactions are not statistically distinguishable.Interactions with each additional prior manipulation exposure were not statistically distinguishable for hireability, trustworthiness, or advancement.Source: Prespecified temporal carryover analysis · p. 22

Evaluator heterogeneity

Downstream AI-generated-versus-real differences vary descriptively by evaluator gender and race, but the subgroup evidence is limited and does not establish stronger responses among men.

Evaluator heterogeneity

  • Male evaluators show statistically distinguishable within-group contrasts on all three outcomes, but nonsignificant treatment-by-gender interactions do not establish a stronger response than among women.For evaluator race, the hireability exposure interaction is significant, including a significant difference between Black and White evaluators; race-by-gender three-way interactions are not significant.Source: Evaluator heterogeneity · p. 22

Small pilot of the AI-evaluator extension

The authors piloted a multimodal-language-model extension to the candidate-evaluation sequence, but did not scale it because models inconsistently used image information in hiring judgments.

Small pilot of the AI-evaluator extension

  • The pilot tested whether images supplied through multimodal APIs could function as candidate information in the Study 2 evaluation sequence.It used deliberately out-of-domain control images.Source: Small pilot of the AI-evaluator… · p. 22
  • Models identified image content when asked directly but did not consistently incorporate visual information into hiring judgments, so the extension was not scaled.Pilot outputs were retained as implementation diagnostics, not substantive evidence.Source: Small pilot of the AI-evaluator… · p. 22

Rendered-attribute adjustment set

The adjustment strategy used nested models: Model 2 added four image-derived controls, and Model 3 added 38 further attribute probabilities while retaining the Model 2 controls.

Rendered-attribute adjustment set

Loading 2610.22553v1…