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

cs.CYOct 15, 2026

Generative AI Produces and Transmits Beauty Bias

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

Organizations: Cornell University, Ithaca, NY, USA · University of Florida, Gainesville, FL, USA · University of Pennsylvania, Philadelphia, PA, USA

Abstract

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.

Figures & tables

Figure 1: Generated faces are generally more attractive than photographs across automated and human evaluations. A , Generated–photographed attractiveness difference for each of 27 image generators, expressed as the mean Cohen’s $d$ across the three automated attractiveness instruments. B , Standardized attractiveness of each prompted demographic category within each generator. Values are standardized within generator and instrument and averaged across the three automated instruments; rows follow the generator ordering in panel A. C , Mean human attractiveness rating (1–7) of AI-generated faces by generator in the human-rated subset of the audit. The panel shows the 23 generators represented by at least 10 rated images; the photographed comparison mean is shown as a reference and error bars indicate 95% confidence intervals around generator image means. D , Probability that a generated face scores above the median professional fashion model, averaged across the two primary automated attractiveness measures (SCUT and MEBeauty). Filled points show all generated faces and open points female-prompted faces. The gray dashed line marks the pooled probability across all generated faces (23%), and the gold dashed line marks the pooled probability among female-prompted generated faces (37%); generators are ordered by the all-face probability in this panel.

Figure 2: Generated attractiveness is racially structured across automated and human evaluations. A , Mean standardized attractiveness of prompted demographic categories across all 27 systems and three automated measures. B , Mean Study 1b human attractiveness ratings for AI-generated and photographed faces within six broad racial and ethnic categories; error bars indicate 95% confidence intervals around image means, and labels give the generated-minus-photographed difference. Categories are ordered by mean human attractiveness rating of AI-generated faces.

Measure	W–B Model 1	W–B Model 2	W–B Model 3	% remains	W–Nig Model 1	W–Nig Model 2	W–Nig Model 3	% remains
SCUT	+0.615	+0.577	+0.472	77%	+0.768	+0.722	+0.593	77%
MEBeauty	+0.744	+0.757	+0.737	99%	+1.002	+0.970	+0.885	88%
CelebA/SwinFace	+0.167	+0.132	0.000	0%	+0.175	+0.135	-0.005	-3%

Table 1: White–Black and White–Nigerian differences in generated-face attractiveness under nested adjustment. Model 1 is unadjusted. Model 2 adds perceived age, skin lightness, adiposity, and predicted blur. Model 3 retains those controls and additionally adds 38 CelebA-style facial-attribute probabilities from SwinFace, excluding Attractive and Blurry. CelebA/SwinFace Model 3 estimates should be read cautiously because its attractiveness score and the added Model 3 attributes come from the same model; a negative percent remaining denotes a slight sign reversal after adjustment.

Figure 3: Exposure to AI-generated faces lowers evaluations of identical real candidates. Points show adjusted contrasts relative to the real-face exposure condition; whiskers show 95% confidence intervals. Hireability and trustworthiness contrasts are shown in rating-point units; advance-to-interview contrasts are shown in log-odds units, so the null value is zero for all three outcomes. Retouched-versus-real estimates overlap the null for all three outcomes, whereas AI-generated-versus-real estimates indicate lower evaluations and exclude the null. Prespecified one-sided Holm-adjusted $P$ values for the AI-generated-versus-real contrasts are 0.037, 0.037, and 0.011 for hireability, advancement, and trustworthiness, respectively.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1: Illustrative examples from the generated-attractiveness distribution. Images were selected algorithmically from fully specified prompts near the upper and lower tails of the three-measure standardized attractiveness summary. Display constraints do not affect the analyses; complete prompt metadata are reported in Table S5 .

Figure S2: Illustrative outputs across successive Midjourney generations. Columns show Midjourney v6, v7, and v8 from left to right; rows hold the prompt category constant. These examples illustrate changes in visual style across releases and are not used for inference.

Figure S3: Human attractiveness ratings by image source and sex in Study 1b. Mean ratings are higher for AI-generated than photographed faces among both female- and male-presenting images in the rated pool.

Figure S4: White–Black attractiveness differences are larger for generated than photographed faces. Bars show the White–Black difference when each attractiveness instrument evaluates photographed faces and generated faces, together with the White–Black difference in each group’s generated-minus-photographed change. The third quantity is standardized within instrument and is not the arithmetic difference between the first two bars.

Generator	Vendor	Year	$d$ (SCUT)	$d$ (MEBeauty)	$d$ (CelebA/SwinFace)	W–B (SCUT)	W–B (MEB)	W–Nig (SCUT)	W–Nig (MEB)
playground-2.5	Playground	2024	+2.25	+2.05	+0.90	+0.55*	+1.27*	+0.76*	+1.69*
sana	NVIDIA	2025	+2.14	+0.10	-0.34	+0.25	+0.48*	+0.34*	+0.58*
sd-xl	StabilityAI	2023	+1.98	+1.53	+0.68	+0.57*	+0.86*	+0.86*	+1.15*
sd-3.5-large	StabilityAI	2024	+1.94	+1.66	+0.93	+0.48*	+0.82*	+0.53*	+0.83*
pixart-sigma	PixArt	2024	+1.84	+1.69	+0.97	+0.58*	+0.73*	+0.87*	+1.07*
flux-schnell	BFL	2024	+1.83	+1.26	+0.89	+0.48*	+0.44	+0.62*	+0.95*

Appendix

Table S1: Generated-versus-photographed attractiveness (Cohen’s

d

) under all three automated measures and White–Black and White–Nigerian differences in generated-face attractiveness under SCUT and MEBeauty, by generator. Asterisks mark gaps significant after Bonferroni correction across all 162 gap tests.

Stage	Outcome	Real mean	Retouched mean	AI-generated mean	Retouched vs. real	AI-generated vs. real	$P_{\mathrm{Holm}}$ (AI vs. real)
Direct	Hireability	5.418	5.455	5.532	$b=.033$ , $P=0.624$	$b=.162$ , $P=0.015$	–
Direct	Advance to interview	82.54%	82.72%	84.80%	OR $=.958$ , $P=0.803$	OR $=1.218$ , $P=0.265$	–
Direct	Trustworthiness	5.278	5.314	5.361	$b=.018$ , $P=0.789$	$b=.130$ , $P=0.047$	–
Downstream	Hireability	5.262	5.256	5.065	$b=-.011$ , $P=0.881$	$b=-.146$ , $P=0.042$	0.037
Downstream	Advance to interview	77.38%	78.23%	72.46%	OR $=1.060$ , $P=0.657$	OR $=.768$ , $P=0.037$	0.037
Downstream	Trustworthiness	5.223	5.229	5.001	$b=-.014$ , $P=0.828$	$b=-.172$ , $P=0.007$	0.011

Appendix

Table S2: Direct and downstream Study 2 evaluations. Direct continuous-outcome coefficients and advancement odds ratios are mixed-effects estimates; downstream continuous outcomes use mixed-effects models and downstream advancement uses GEE.

Temporal test	Outcome	Retouched vs. real	AI-generated vs. real
Cumulative prior manipulation count	Hireability	$-.017$ ( $P=0.307$ )	$-.017$ ( $P=0.299$ )
	Trustworthiness	$-.005$ ( $P=0.734$ )	$-.002$ ( $P=0.888$ )
	Advance to interview	OR $=1.040$ ( $P=0.255$ )	OR $=1.008$ ( $P=0.813$ )
Immediately follows manipulation	Hireability	$-.068$ ( $P=0.306$ )	$+.083$ ( $P=0.212$ )
	Trustworthiness	$-.040$ ( $P=0.476$ )	$+.082$ ( $P=0.148$ )
	Advance to interview	OR $=.873$ ( $P=0.304$ )	OR $=1.105$ ( $P=0.440$ )

Appendix

Table S3: Temporal interaction estimates in the full

N=570

fielded sample. Entries are treatment-by-temporal-variable interaction estimates relative to unmodified-real exposure. Continuous outcomes are in raw 1–7 rating units; advancement entries are odds ratios.

Condition	Construction	Prior human attractiveness	Calibration check
Real	Eight unmodified photographed identities	Mean $\approx 2.99$ / 7	Reference
Retouched	Same eight identities; calibrated smoothing and brightening	Not independently rated	See retouching calibration section
AI-generated	Eight contemporary generated faces selected for high attractiveness	Mean $\approx 5.75$ / 7	Selected separately for high attractiveness

Appendix

Table S4: Calibration of the three Study 2 manipulation sets. Human ratings are prior Study 1b ratings for the exact unmodified-real and AI-generated fielded images. No independent human attractiveness rating was collected for the retouched twins.

Tail	Generator	Demographic prompt	Sex prompt	Occupation	Mean $z$
Upper	Midjourney v7	Brazilian	female	retail worker	+1.36
Upper	Midjourney v6	East Asian	female	lawyer	+1.36
Upper	Playground 2.5	Mexican	male	pharmacist	+1.33
Upper	SD-XL	Middle Eastern/North African	male	veterinarian	+1.41
Lower	Midjourney v8	Ethiopian	female	construction worker	-1.31
Lower	Midjourney v6	Mexican	female	retail worker	-1.33

Appendix

Table S5: Prompt metadata for the eight images displayed in Figure S1.

Measure	W–B among men	W–B among women	W–B high prestige	W–B low prestige	Female–male
SCUT	+0.648	+0.582	+0.601	+0.629	-0.080
MEBeauty	+0.856	+0.629	+0.696	+0.792	+0.901
CelebA/SwinFace	+0.135	+0.200	+0.188	+0.147	+0.095

Appendix

Table S6: Raw-score differences in the White–Black prompted gap within social context and the overall female–male difference in rendered attractiveness, by measure. High prestige comprises doctor, lawyer, veterinarian, and pharmacist.

Prompt slot	Terms
Race/ethnicity (6)	White, Black, Hispanic/Latino, East Asian, South Asian, Middle Eastern/North African
Nationality (8)	American, Mexican, Brazilian, Nigerian, Ethiopian, German, Indian, Korean
Sex (2)	female, male
Higher-prestige occupation (4)	doctor, lawyer, veterinarian, pharmacist
Lower-prestige occupation (4)	cook, construction worker, nurse, retail worker
Default conditions (3)	demographic, sex, or occupation term omitted

Appendix

Table S7: Prompt design for the generated-image audit. Every fully specified combination of demographic term, sex term, and occupation was issued to every system five times. Additional conditions omitted the demographic, sex, or occupation term.

Measure	Real W–B gap	Generated W–B gap	W–B difference in generated–photographed change
SCUT-trained	+0.23	+0.61	+0.73
MEBeauty-trained	+0.38	+0.74	+0.69
CelebA/SwinFace	+0.03	+0.17	+0.45

Appendix

Table S8: White–Black attractiveness differences in photographed and generated faces, and the White–Black difference in the generated–photographed change relative to group-specific photographed baselines. Positive values indicate higher scores for White relative to Black faces. The first two columns report differences on each instrument’s native score scale; the White–Black difference in generated–photographed change is expressed in standardized units and is therefore not the arithmetic difference between those columns.

Additional Model 3 attribute probabilities
swin_chubby	swin_heavy_makeup	swin_gender
swin_oval_face	swin_pale_skin	swin_smiling
swin_young	swin_bald	swin_bangs
swin_black_hair	swin_blond_hair	swin_brown_hair
swin_gray_hair	swin_receding_hairline	swin_straight_hair
swin_wavy_hair	swin_wearing_hat	swin_arched_eyebrows

Appendix

Table S9: Exact additional SwinFace/CelebA-style attribute probabilities entered in Model 3. Model 3 also retains all four Model 2 controls: DeepFace age, STONE-derived skin lightness, CLIP-based facial adiposity, and SwinFace Blurry. Attractive, Blurry, SwinFace age, and SwinFace expression outputs are not part of this additional 38-variable block.

Photographed source	SCUT $d$ [95% CI]	MEBeauty $d$ [95% CI]	CelebA/SwinFace $d$ [95% CI]
FairFace	1.370 [1.129, 1.610]	1.134 [0.799, 1.444]	1.262 [0.889, 1.621]
IdenProf	1.823 [1.554, 2.095]	0.802 [0.382, 1.189]	1.000 [0.718, 1.259]
SCUT-FBP5500	0.479 [0.314, 0.639]	0.176 [-0.105, 0.434]	-0.980 [-1.092, -0.873]
MEBeauty	0.062 [-0.153, 0.268]	0.017 [-0.177, 0.193]	-0.782 [-0.923, -0.650]
Chicago + extensions	1.975 [1.705, 2.238]	1.075 [0.820, 1.310]	0.389 [0.194, 0.574]
London Set	1.517 [1.256, 1.779]	0.702 [0.449, 0.935]	-0.215 [-0.604, 0.135]

Appendix

Table S10: Source-specific robustness of the generated–photographed attractiveness difference. For each photographed source and instrument, Cohen’s

d

was calculated separately for each of the 27 generators and then averaged across generators. For the pooled baseline, the six photographed sources were first combined at the image level and each generator was compared with that pooled distribution; the pooled value is therefore not a sample-size-weighted average of the six source-specific effect sizes. Brackets report 95% bootstrap confidence intervals obtained by resampling generator-level effect sizes. Positive values indicate higher attractiveness scores for generated faces.

Generator	Rated images	Human mean	Difference vs. photos	95% CI	Holm-adjusted $P$
GPT Image 1.5	37	5.16	+2.06	[1.84, 2.29]	$<0.001$
FLUX.1 dev	29	5.03	+1.93	[1.60, 2.26]	$<0.001$
GPT Image 2	29	4.91	+1.81	[1.52, 2.10]	$<0.001$
Midjourney v8	59	4.72	+1.62	[1.37, 1.87]	$<0.001$
Midjourney v7	20	4.67	+1.57	[1.25, 1.89]	$<0.001$
Lumina-2	29	4.64	+1.54	[1.22, 1.87]	$<0.001$

Appendix

Table S11: Generator-level human attractiveness ratings relative to photographed faces in Study 1b. The photographed-image mean is 3.10 (614 images). Each row compares image-mean ratings for one generator with the photographed pool using Welch’s

t

-test;

P

values are Holm-adjusted across the 23 generator-level comparisons. Only generators represented by at least 10 rated images are included, matching Fig. 1C.

Gender
Characteristic	Study 1b human ratings	Study 2 hiring experiment
Participants	550	570
Demographic responses	549	570
Woman	276 (50.3%)	315 (55.3%)
Man	268 (48.8%)	248 (43.5%)
Non-binary	1 (0.2%)	3 (0.5%)

Appendix

Table S12: Realized participant characteristics for the two human studies. Study 1b enrolled 550 participants; demographic responses were available for 549. Study 2 enrolled 570 participants. Percentages use the number with demographic data as the denominator. Study 1b’s Asian category combines East/Southeast Asian and South Asian responses.

Demographic Bias in T2I Models

Demographic bias in text-to-image models is the unequal or stereotyped representation of people across generated images.

See all 31 papers

Attractive Faces Skew Text-to-Image Reward Models and Leaderboards
Oct 20, 2026Harvey Mannering, Zhiwu Huang, Adam Prügel-BennettT2I Generation Evaluation Human Preference Evaluation
University of Southampton
Gender Bias in Generative AI-assisted Recruitment Processes
Mar 12, 2026Martina Ullasci, Marco Rondina, Riccardo Coppola +1AI and Labor Gender Bias in Language Models
Department of Control and Computer Engineering Politecnico di Torino, Italy Turin, Italy
Who Defines Fairness? Target-Based Prompting for Demographic Representation in Generative Models
Apr 22, 2026Marzia Binta Nizam, James DavisSocial Bias in Language Models Algorithmic Fairness
University of California, Santa Cruz, USA

Oct 20, 2026J/K move · Enter open · S save

Attractive Faces Skew Text-to-Image Reward Models and Leaderboards

Harvey Mannering, Zhiwu Huang, Adam Prügel-Bennett

University of Southampton

T2I Generation Evaluation Human Preference Evaluation Human Preference Modeling Preference Optimization for Diffusion Models

Open paper AskPDF arXiv

Human preference is now central to the evaluation, guidance, finetuning, and ranking of text-to-image models. However, the susceptibility of human preferences to bias in image generation remains underexplored. We identify a key observation: for human faces, perceived attractiveness overrides image quality in preference judgments, even when comparing real photographs of comparable fidelity. In a controlled user study, annotators rank more attractive faces as higher quality, even when unattractive real faces are compared with attractive AI-generated ones. We examine the downstream consequences of this bias. Firstly, simulating Elo rankings from user study data shows attractiveness can distort a leaderboard's ranking, elevating synthetic attractive faces over real, unattractive ones. Secondly, the same bias reappears in the human preference models. Editing 40 annotated facial attributes in isolation, we find that attractiveness is the strongest correlate of score for four of five human preference models (PickScore, MPS, HPSv2.1, HPSv3), and the strongest predictor of score in regression analysis. We further show that preference aligned generation shifts faces toward younger, more feminine faces, and can override explicit prompts for unattractive faces, thereby limiting user control. Finally, we explore mitigation strategies, both for data collection and post training on human preference models. These findings highlight the need for preference models that disentangle image quality from human cognitive bias, enabling more reliable evaluation.

Abstract

Demographic Bias in T2I Models

Explore similar work

Attractive Faces Skew Text-to-Image Reward Models and Leaderboards

Gender Bias in Generative AI-assisted Recruitment Processes

Who Defines Fairness? Target-Based Prompting for Demographic Representation in Generative Models