Underweight and Overrepresented: Body Size Bias in AI-Generated Depictions of Women

Underweight and Overrepresented: Body Size Bias in AI-Generated Depictions of Women | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Underweight and Overrepresented: Body Size Bias in AI-Generated Depictions of Women Justin Thomas, Fahad Al-Beyahi, Najud Al-Dugheithir, Mohammad Amin Kuhail This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9266968/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The bias towards portraying thinner women in traditional media has been well-documented. However, there is limited research on such bias in images generated by artificial intelligence (AI). This study examined viewers' perceptions of AI-generated depictions of women, assessing whether these images were perceived as underweight, normal weight, or overweight. Using standardised prompts that did not specify body shape or size, images of women across various occupations (e.g., “draw a female lawyer”) were created using three popular generative AI tools. Participants ( N =129), adult women from the US, rated the body size of the AI-generated images using a shortened version of the Stunkard figure rating scale. Overall, the proportion of images perceived as underweight was more than threefold the expected rate of underweight among women in the US population. Across all platforms, images of women with greater occupational prestige were more likely to be rated as underweight. The findings suggest a systematic body size bias in text-to-image AI with implications for population-level body image dissatisfaction and eating disorders. Psychology Artificial Intelligence and Machine Learning artificial intelligence body image bias eating disorder online media Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction The launch of OpenAI's ChatGPT in November 2022 introduced the internet-using public to generative AI. Within eight weeks of the platform’s launch, it had gained 100 million users (mostly in the USA), vastly surpassing the growth rates of all earlier online platforms, making it the fastest-growing consumer software application in history(Thomas et al., 2025). Subsequently, other companies launched similar services (Google Gemini, MetaAI assistant, Deepseek), and several began offering text-to-image features, allowing users to prompt AI to generate images from natural-language inputs. These text-to-image AI tools are rapidly changing how people source and produce images for various purposes, from personal enjoyment to building visual brand identities. According to market intelligence reports, around 34 million such images are generated each day, and it is estimated that since 2022, over 15 billion AI images have been created (DigitalSilk, 2026). Given the rapid proliferation of AI tools and continuous improvements in their output quality, AI-generated images are likely to become a major part of the visual environments people encounter daily (Messingschlager & Appel, 2025). One of the challenges of such a scenario is that AI can introduce or reinforce existing social biases at scale. Such algorithmic bias, like bias in general, typically involves an inclination toward preferencing one individual or group over another (The Oxford English dictionary, 2026) or, in the more abstract sense the “..deviation of a measured or calculated quantity from its actual (true) value, such that the measurement or calculation is unrepresentative of the item of interest( APA Dictionary of Psychology , 2007). For example, if the actual (true) rate of female chief executive officers (CEOs) was 12% and our image-to-text algorithm, when repeatedly prompted to “draw a CEO”, generated female CEOs at a rate of 0.12%, we could describe this as representational gender bias. The origins of such biases within AI algorithms are usually linked to unrepresentative training data, meaning the data used to train the system fails to adequately reflect the diversity of the real-world population it aims to serve. Furthermore, the designers of such systems may inadvertently embed their own biases and assumptions within their creations(Noble & Roberts, 2019). Researchers exploring biases in AI text-to-image systems have begun documenting cases. In one study, nine different text-to-image platforms were used to visually represent professionals working in the prestigious fields of law, medicine, engineering, and scientific research(Gorska & Jemielniak, 2023). Based on viewer categorisation of the resultant output, 76% of the images were deemed male, while only 8% were judged female. The bias persisted across all professions, but the extent of the gender skew varied greatly by AI platforms. Another study focused specifically on the representation of occupations in science, technology, engineering, and maths (STEM) fields. This study asked Midjourney (an AI service specialised in image generation) to “Imagine a portrait of a [insert scientist/person in science],” where the inserted word was drawn from a list of STEM occupations, e.g., biologist, physicist. Of the 84 images generated, all were of white males, suggesting a particularly extreme representational bias against women and people of colour(Messingschlager & Appel, 2025). Even when the AI was explicitly prompted to depict female scientists, these were typically portrayed as much younger than their male counterparts. In this context, youth might be interpreted as “less experienced”. The bias towards portraying females as younger than their male counterparts was also observed in a study of 1.4 million digital images(Guilbeault et al., 2025). Despite no systematic age differences between women and men in the workforce according to the US Census, the study found that women were portrayed as younger than men across a wide range of occupations, with the bias being even more evident in higher-status roles. Another study examined the representation of gender and ethnic markers in AI-generated imagery across occupations using three text-to-image platforms, discovering that both ethnicity and gender were over- or underrepresented to varying extents, thereby reinforcing existing social stereotypes(Luccioni et al., 2023). For instance, all three platforms typically depicted dentists as male (83.57% - 62.50%), whereas dental assistants were almost exclusively portrayed as females (96.79% - 94.34%). A similar trend was observed for doctors (98.09%–87.14% male) and nurses (100%–98.06% female). Exploring the potential bias- and stereotype-amplifying implications of text-to-image AI in the context of global health imagery, one study sought to craft prompts that would challenge prevalent global health tropes/stereotypes, such as the so-called “white saviour” trope(Alenichev et al., 2023). To this end, Midjourney was tasked with depicting images of Black African doctors attending to sick White children. Despite the use of prompts crafted to challenge prevalent stereotypes, the AI invariably, across 300 images, rendered the child patients as Black, even though the prompts explicitly specified White children. Similarly, Black African doctors were sometimes depicted as White, despite that being the stereotype the researchers had sought to invert (Alenichev et al., 2023). The Authors concluded that “..despite the supposed enormous generative power of AI, it proved incapable of avoiding the perpetuation of existing inequality and prejudice.” (p. 1496). Almost invariably, the existing literature on bias in AI imagery warns of its potential to perpetuate stereotypes and further entrench social biases, inequalities, and power imbalances (Alenichev et al., 2023; Guilbeault et al., 2025; Szymański et al., 2025). Whether intentional or inadvertent, biases in AI-generated imagery can impact public perceptions, shape opinions, and affect the assumptions individuals make about others or themselves. To date, however, much of this research has remained focused on gender and ethnicity-related biases. We are unable to find any studies that have systematically examined how AI’s text-to-image functions represent body shape and size, which can also serve as a rough approximation of weight status(Parzer et al., 2021). This is an important focus given the extensive and widely reported links between media imagery, body image dissatisfaction and eating disorders, especially (although not exclusively) among females. Since the 1990s, eating disorders have become a significant and increasingly global public health problem (Austin, 2012; Gordon, 2000). Those experiencing such disorders demonstrate heightened mortality rates, with the risk of death more than three times higher than that of the general population, while among those diagnosed with anorexia nervosa, the mortality rate rises to fivefold (Krug et al., 2025). The onset and maintenance of eating disorders such as anorexia and bulimia nervosa is multifaceted, involving biological, psychological and socio-cultural factors(American Psychiatric Association, 2022). Within the context of sociocultural influences, however, extensive research indicates that the promotion of increasingly thin and unattainable female body image ideals in the media can play a significant role in the onset of such disorders(Gordon, 1990, 2001; Groesz et al., 2002). The proposed mechanism of action involves self-comparison with media-promoted ideals, which fosters body image dissatisfaction (Laker & Waller, 2022) and maladaptive body image attitudes, such as overvaluing thinness(Fairburn et al., 2003). In turn, body image dissatisfaction and maladaptive body image attitudes can give rise to maladaptive weight-control measures (Larson et al., 2021) and eating disorders such as anorexia and bulimia nervosa(American Psychiatric Association, 2013). The idea of bias toward portraying thinner female body image ideals in traditional media is well-documented. A seminal study by Garner et al. (1980) reported decreases in the weight and body dimensions of US beauty pageant contestants between 1959 and 1978, coinciding with rising weight norms for US women. More recently, Maymone et al (2019) reported the same patterns for Victoria’s Secret models between 1995 and 2018. This bias toward celebrating thinner female body images, typically depicted in the media, highlights discrepancies between society-wide body image ideals and anthropometric realities(Thompson et al., 1999). The current rate of overweight and obesity among US women stands at 70.3% [67.1 – 73.3], according to the World Health Organisation(2022), with similar rates reported by the US National Centre for Health Statistics(Fryar et al., 2026). Discrepancies between body image aspirations/ideals and anthropometric realities or perceptions are central to most definitions of body image dissatisfaction(Cash & Szymanski, 1995; Garner, 1997; Khalaf et al., 2011; Thompson & Altabe, 1991; Thompson et al., 1999). AI’s text-to-image function threatens to further amplify the existing body image biases, further exacerbating societal levels of body image dissatisfaction and inadvertently promoting eating disorders. In this study, we aim to explore the extent to which text-to-image AI depicts women as either underweight, normal weight, or overweight. Given that AI systems are typically trained on images and captions scraped from the public internet, including digitised legacy media(Whang et al., 2023), we make the following hypotheses. Hypothesis 1: The rate of perceiving AI-depicted women as underweight will be higher than the estimated age-adjusted prevalence of underweight for women in the US population. Hypothesis 2: The rate of perceiving AI-depicted women as underweight will be higher than the estimated age-adjusted prevalence of underweight for women in the US population. Furthermore, considering previously observed platform-level differences in gender bias, we also propose the following hypothesis. Hypothesis 3: Perceived body size of AI-depicted women will vary by platform. Finally, previous studies have documented increased levels of gender (male) and age (young female) bias linked with higher occupational status. The same might be true for thinness. We suspect an algorithmic variant of the “halo effect” (Thorndike, 1920) is at work, whereby socially desirable traits such as thinness, youthfulness, and occupational prestige are combined for females. Based on these ideas, we formulate the following hypothesis. Hypothesis 4: AI-depicted women in roles with higher occupational prestige will be rated as having thinner body sizes. Methods Participants Participants were recruited through Prolific, a crowdsourcing platform dedicated to social science research. Prolific is among the most widely used recruitment platforms for research in this field and has been reported to deliver higher-quality data than similar services (Peer et al., 2017; Stritch et al., 2025). The present non-probability quota sample consisted of 129 young women (aged 18–34) living in the United States. The focus on young women was motivated by previous research showing that this age group and gender are particularly vulnerable to body dissatisfaction and engaging in body-focused social comparison online, that is, comparing one's physical appearance with digital images of others (Merino et al., 2024). The focus on US residents was justified by the relatively widespread use of AI in US society and by the fact that the AI platforms under examination are founded and hosted in the US. Participants were selected from a pool of 22,689 eligible individuals and were allowed to participate only on a personal computer or tablet. In addition to age, they were asked to report their height, weight, and current employment status. Most (73.43%) were employed, while 10.15% were engaged in education, 14.06% were unemployed, and 2.34% identified their employment status as other. Table 1 provides details of all other participant characteristics. Table 1 Sample Characteristics Age BMI Mean 26.32 26.95 SD 3.05 7.89 Median 26.50 26.95 IQR 5 9.65 Min 19 16.73 Max 32 58.58 Measures and Materials Stunkard Figure Rating Scale (FRS) The original FRS (Stunkard et al., 1983) consists of nine silhouette drawings of body shapes on a gradient from underweight to obese, with separate figure sets for males and females. Previous studies report good convergent and discriminant validity for the scale, as well as test-retest reliability, across diverse populations (Lo et al., 2011; Scagliusi et al., 2006). Each figure is assigned a number/score, 1 to 9; higher scores indicate larger body sizes. Simple to administer, this measure has been widely used to assess one's current body size, one’s body size aspirations, and to assign perceived body size or weight status to others. For example, Cardinal et al. (2006) used the FRS to assign figure ratings to adult women viewed on videotape. In this instance, observer ratings were correlated with the actual BMIs of the women being rated, r [70] = .91. Similarly, in the present study, rather than using the scale for own-body evaluation, participants were asked to rate the body size of AI-generated depictions of women. The current study used an abbreviated six-silhouette version of the FRS, which facilitated a more parsimonious categorisation of the AI-generated depictions as either underweight, normal weight, or overweight. Based on Parzer et al. (2021), a large population study exploring the correspondence between FRS silhouettes and clinically assessed BMI, silhouette 1 was taken to represent underweight, silhouettes 2, 3, and 4 represented normal weight, and silhouettes 5 and 6 were categorised as overweight. Figure 1 details the estimated correspondence between the silhouettes, BMI, and weight status. AI-Generated Depictions of Women We chose three general-purpose AI platforms/assistants for image generation based on their current popularity, free access and their provision of text-to-image features. These platforms were ChatGPT 5.3, Google AI Studio and Meta AI assistant. Including additional platforms was beyond the scope of this initial investigation. To generate images, we used a standardised natural language prompt across all platforms: “Draw a female (insert occupation – e.g., firefighter), head to toe, facing the viewer”. In the few cases when the AI hallucinated, for example, depicting a woman with three legs, it was instructed to “try again”. There were only three instances where this additional step was required. All images were generated between 1 st and 14 th September 2025. Occupational Prestige The AI-generated images spanned three categories: high, middle and low occupational prestige. This categorisation was based on data from the National Opinion Research Centre (NORC), in which occupational categories listed in the 1980 US Census were rated on a 0-100 prestige scale by approximately 1,200 independent raters (Hauser & Warren, 1997; Nakao & Treas, 1992). This list contains 740 occupations, and we randomly selected 12 occupations from the upper (Lawyer, CEO), middle (Librarian, Chef), and lower (e.g., bus driver, seamstress) terciles of the occupational prestige list. We excluded any occupations in which body shape and size can be associated with the role, for example, athletes and cabin crew. There were 36 images in total, equally distributed across the occupational prestige categories and AI platforms (3 x 12). Figure 2 provides examples of images used in the study from each occupational prestige category. Procedure The study protocol was conducted online and received prospective ethics approval from the institution's internal review board. Participants received a brief overview of the study and were informed that the data collection would remain anonymous. They then could indicate their consent to participate by clicking 'continue'. After answering demographic questions, participants completed the body size rating task, which involved 36 trials in which they used the FRS to estimate an approximate body size for each AI-generated image (see figure 3). The presentation order of AI-generated images was arranged using a permutation that prevented consecutive repetition of occupational prestige levels or image sources (i.e., AI platforms). All data were collected on 9 November 2025. The average task completion time was 7 minutes 43 seconds, with all participants completing every trial. Results In this study, 16.64% of the AI-generated depictions of women were rated as overweight (body image figures 5-6), while 7.76% were perceived as being underweight (Silhouette 1). By comparison, based on age-standardised estimates from the World Health Organisation(2022), 70.3% [67.1 – 73.3] of US females are overweight (BMI >= 25), while the age-standardised rate of underweight (BMI < 18) among US females is 2.3% [1.5 – 3.3]. Figure 4 below visually compares the percentages of AI-generated female images rated as underweight, normal weight, and overweight with US weight-status norms reported by the World Health Organisation (WHO) in 2022. These findings lend support for hypotheses 1 and 2. Suggesting that, in AI portrayals of women, underweight is perceived more frequently than in the general population, and conversely, overweight is perceived less frequently than in the general population. The image most frequently classified as underweight, that is, the image that most viewers rated as matching the FRS silhouette 1, was image 7, see figure 5 below. Previous research suggests that this body-silhouette corresponds to a BMI of 17.88 (Parzer et al., 2021). By contrast, the image most frequently classified as underweight, that is, the image that most viewers rated as matching the FRS silhouette 5 or 6, was image 24, see figure 5 below. Previous research suggests that this body silhouette corresponds to a BMI of 25.68–28.75 (Parzer et al., 2021). AI body-size ratings were uncorrelated with participants' self-reported BMI; however, younger participants tended to provide higher (larger) body-size ratings, r [127] = .244, p = .005. Supportive of hypothesis 3, the body size ratings for the AI-generated depictions of women varied widely across platforms, with the frequency of underweight ratings lowest on ChatGPT, which also received the highest frequency of overweight ratings for its generated images. The percentage of underweight, normal weight, and overweight by platform and occupational prestige is detailed in Figure 6 The mean figure rating score for the ChatGPT-generated images was 4.10 ( SD = 1.19), for GoogleAI it was 2.65 ( SD = 0.95) and for MetaAI it was 2.78 ( SD = 1.15). A repeated-measures ANOVA was performed to evaluate the effect of the AI platform on body size ratings. Mauchly’s test indicated that the assumption of sphericity was violated, χ²(2) = 125.59, p < .001. Degrees of freedom were corrected using Greenhouse-Geisser estimates (ε = .92). The effect of the AI platform on body size ratings was significant, F (2, 1535) = 915.98, p < .001, partial η² = .37. Post hoc pairwise comparisons indicated that GoogleAI images were perceived as thinner than those generated by MetaAI ( p < .001) and ChatGPT ( p < .001). Supporting hypotheses 3 and 4, Figure 7 plots the means of body size ratings across the three AI platforms and occupational prestige. The mean figure rating score for the AI-generated images of women working in high-status occupations was 2.66 ( SD = 1.15); for middle-status occupations, 3.12 (SD = 1.16); and for low-status occupations, 3.758 ( SD = 1.29). A repeated-measures ANOVA was performed to evaluate the effect of occupational status on perceptions of body size. Mauchly’s test indicated that the assumption of sphericity was violated, χ²(2) = 95.48, p < .001. Degrees of freedom were corrected using Greenhouse-Geisser estimates (ε = .94). The effect of the AI platform on body size ratings was significant, F (2, 1535) = 325.98, p < .001, partial η² = .17. Post hoc pairwise comparisons indicated that images of women in high status occupations were perceived as thinner than those in middle status occupations ( p < .001) and low status occupations ( p < .001). Discussion This study examined the perceived body size and, by extension, perceived weight status of AI-generated images of women. As hypothesised, the proportion of perceived underweight was much higher (more than three times greater) than that reported in the general adult female population of the USA. Similarly, the proportion of AI-generated images of women perceived as overweight was substantially lower than US population estimates. This pattern was even more pronounced for depictions of women working in more prestigious occupations. A quantitative analysis of the figure ratings for the AI images showed that portrayals of women in high-prestige roles (e.g., CEO, Lawyer, Architect) were rated as significantly thinner than those in roles associated with middle and lower levels of occupational prestige (e.g., Teacher, Garbage Collector). The links with occupation prestige may reflect the tendency for those of lower socioeconomic status to have higher rates of overweight and obesity in Western societies (Silventoinen et al., 2024). However, it might also reflect a clustering of socially desirable attributes (youth, beauty, occupational prestige, and thinness) akin to the halo effect(Thorndike, 1920), in which the perception of one positive (socially desirable) trait leads to unfounded inferences about the presence of other positive traits. At least one study suggests AI can be prone to such halo effects (Gulati et al., 2025), and the idea is certainly worthy of future research. Overall, these findings support the hypothesis that text-to-image AI systematically overrepresents very thin (potentially underweight) women while underrepresenting overweight women, at least in comparison to current US weight norms. This body-size-related representational skew extends the list of biases (gender, age, and ethnicity) identified in earlier studies of AI’s text-to-image function (Alenichev et al., 2023; Guilbeault et al., 2025; Szymański et al., 2025). The fact that the three platforms differed significantly in the extent to which they produced body-size bias suggests that engineering solutions could minimise or even eliminate such bias. These platform-level differences have also been observed in relation to the magnitude of gender-biased portrayals (Gorska & Jemielniak, 2023), again suggesting that such biases are, to some extent, platform-specific and not beyond attenuation. Attempts at bias correction, however, raise questions about what the optimum outputs should be: a true representation of reality, engineered equality, or overrepresentation of historically marginalised groups? From a public health perspective, we argue that, in the context of body size, the optimal default would be a body size that reflects a healthy weight status, avoiding the extremes of underweight and overweight. If such extreme images are desired, then users can add additional prompts. Currently, at least two of the AI platforms are generating images of women that are widely perceived as underweight, at rates far exceeding population-level estimates of this weight status. It appears likely that Western society’s promotion of thinner, more tubular female body ideals since the 1960s (Bell, 1985; Bruch, 1974; Gordon, 2000) is well reflected in the legacy data used to train current AI systems. However, explicit design or aesthetic choices made by those developing such systems might also contribute to overrepresenting women broadly perceived as underweight. For instance, analysis of female video game characters indicates that technology developers often intentionally portray these characters with ultra-thin, unrealistic body types (De la Torre-Sierra & Guichot-Reina, 2025). Regardless of how this body-size bias has been established in text-to-image AI systems, it seems likely that, if left unaddressed, it will perpetuate and possibly even amplify unrealistic and unhealthy body image ideals. As the use of AI text-to-image systems grows exponentially, we may inadvertently promote body image dissatisfaction, upward social comparison(Laker & Waller, 2022), and the “overvaluing of thinness”(Fairburn et al., 2003), all of which appear to increase the risk of eating disorders among vulnerable individuals(American Psychiatric Association, 2022). If left unchecked, we risk inadvertently promoting anorexigenic AI. This is an important and emerging concern that has, to date, received limited research attention. This study provides a preliminary examination of the default female body sizes generated by AI text-to-image functionality on three widely used platforms. The study has several important limitations. Firstly, the exploratory nature of this study precluded the inclusion of more platforms, more images, and a larger sample size. Although adequate for the current study, future explorations of this question would benefit from larger samples and more images drawn from a broader array of platforms. Similarly, the current study’s exclusive focus on women limits generalizability and the ability to contrast and compare with portrayals of male body size. Additionally, the use of a crude silhouette figure rating scale (FRS) limits measurement sensitivity and accuracy. For instance, the FRS has only one silhouette corresponding to underweight, whereas there are multiple silhouettes for overweight and normal weight. Future explorations of body-size bias in AI imagery should extend to males, and more sophisticated, sensitive measures should be used in the image-rating process. Setting limitations aside, the present study reveals body-size bias in AI-generated portrayals of women. These insights can inform health policies targeting the digital determinants of health, support the efforts of eating disorder advocacy groups, and assist those in the technology sector in monitoring and promoting ethical and safe AI practices. References Alenichev, A., Kingori, P., & Grietens, K. P. (2023). Reflections before the storm: the AI reproduction of biased imagery in global health visuals. The Lancet Global Health , 11 (10), e1496-e1498. https://doi.org/10.1016/S2214-109X(23)00329-7 American Psychiatric Association. (2013). Diagnostic and statistical manual of mental disorders : DSM-5 . American Psychiatric Association. American Psychiatric Association. (2022). Diagnostic and statistical manual of mental disorders : DSM-5-TR . American Psychiatric Association. APA Dictionary of Psychology . (2007). American Psychological Association. Austin, S. B. (2012). A public health approach to eating disorders prevention: It’s time for public health professionals to take a seat at the table. BMC Public Health , 12 (1), 854. https://doi.org/10.1186/1471-2458-12-854 Bell, R. M. (1985). Holy Anorexia . University of Chicago Press. Bruch, H. (1974). Eating Disorders . Routledge & Kegan Paul. Cardinal, T. M., Kaciroti, N., & Lumeng, J. C. (2006). The figure rating scale as an index of weight status of women on videotape. Obesity (Silver Spring) , 14 (12), 2132-2135. https://doi.org/10.1038/oby.2006.249 Cash, T. F., & Szymanski, M. L. (1995). The Development and Validation of the Body-Image Ideals Questionnaire. Journal of Personality Assessment , 64 (3). https://doi.org/10.1207/s15327752jpa6403_6 De la Torre-Sierra, A. M., & Guichot-Reina, V. (2025). Women in Video Games: An Analysis of the Biased Representation of Female Characters in Current Video Games. Sexuality & Culture , 29 (2), 532-560. https://doi.org/10.1007/s12119-024-10286-0 Fairburn, C. G., Cooper, Z., & Shafran, R. (2003). Cognitive behaviour therapy for eating disorders: a "transdiagnostic" theory and treatment. Behav Res Ther , 41 (5), 509-528. https://doi.org/10.1016/s0005-7967(02)00088-8 Fryar, C. D., Afful, J., & Saif, N. (2026). Health E-Stat 111: Prevalence of Overweight, Obesity, and Severe Obesity Among Adults Age 20 and Older: United States, 1960–1962 Through August 2021–August 2023 (National Health and Nutrition Examination Survey, Issue. https://stacks.cdc.gov/view/cdc/174643 Garner, D. M. (1997). Body image: an in-depth look at how we see ourselves. Psychology Today , (01 Febuary 1997). Gordon, R. A. (1990). Anorexia and Bulimia . Blackwell. Gordon, R. A. (2000). Eating Disorders: Anatomy of a Social Epidemic (2nd ed.). Blackwell. Gordon, R. A. (2001). Eating disorders East and West: A culture-bound syndrome unbound. In M. Nasser, M. A. Katzman, & R. A. Gordon (Eds.), Eating Disorders and Cultures in Transition . Brunner-Routledge. Gorska, A. M., & Jemielniak, D. (2023). The invisible women: uncovering gender bias in AI-generated images of professionals. Feminist Media Studies , 23 (8), 4370-4375. https://doi.org/10.1080/14680777.2023.2263659 Groesz, L. M., Levine, M. P., & Murnen, S. K. (2002). The effect of experimental presentation of thin media images on body satisfaction: a meta-analytic review. Int J Eat Disord , 31 (1), 1-16. https://doi.org/10.1002/eat.10005 Guilbeault, D., Delecourt, S., & Desikan, B. S. (2025). Age and gender distortion in online media and large language models. Nature , 646 (8087), 1129-1137. https://doi.org/10.1038/s41586-025-09581-z Gulati, A., D'Incà, M., Sebe, N., Lepri, B., & Oliver, N. (2025). Beauty and the Bias: Exploring the Impact of Attractiveness on Multimodal Large Language Models . https://doi.org/10.48550/arXiv.2504.16104 Hauser, R. M., & Warren, J. R. (1997). 4. Socioeconomic Indexes for Occupations: A Review, Update, and Critique. Sociological Methodology , 27 (1), 177-298. https://doi.org/10.1111/1467-9531.271028 Khalaf, A., Westergren, A., Berggren, V., & Al-Hazzaa, H. (2011). PP201-SUN Discrepancies between the actual, perceived and ideal body image among female university students in south western Saudi Arabia. Clinical Nutrition Supplements , 6 , 99-100. Krug, I., Liu, S., Portingale, J., Croce, S., Dar, B., Obleada, K.,…Fuller-Tyszkiewicz, M. (2025). A meta-analysis of mortality rates in eating disorders: An update of the literature from 2010 to 2024. Clinical Psychology Review , 116 , 102547. https://doi.org/https://doi.org/10.1016/j.cpr.2025.102547 Laker, V., & Waller, G. (2022). Does comparison of self with others influence body image among adult women? An experimental study in naturalistic settings. Eat Weight Disord , 27 (2), 597-604. https://doi.org/10.1007/s40519-021-01196-3 Larson, N., Loth, K. A., Eisenberg, M. E., Hazzard, V. M., & Neumark-Sztainer, D. (2021). Body dissatisfaction and disordered eating are prevalent problems among U.S. young people from diverse socioeconomic backgrounds: Findings from the EAT 2010-2018 study. Eat Behav , 42 , 101535. https://doi.org/10.1016/j.eatbeh.2021.101535 Lo, W. S., Ho, S. Y., Wong, B. Y., Mak, K. K., & Lam, T. H. (2011). Validity and test-retest reliability in assessing current body size with figure drawings in Chinese adolescents. Int J Pediatr Obes , 6 (2-2), e107-113. https://doi.org/10.3109/17477166.2011.575146 Luccioni, A. S., Akiki, C., Mitchell, M., & Jernite, Y. (2023). Stable bias: evaluating societal representations in diffusion models Proceedings of the 37th International Conference on Neural Information Processing Systems, New Orleans, LA, USA. Maymone, M. B., Laughter, M., Anderson, J. B., Secemsky, E. A., & Vashi, N. A. (2019). Unattainable Standards of Beauty: Temporal Trends of Victoria’s Secret Models from 1995 to 2018. Aesthetic Surgery Journal , 40 (2), NP72-NP76. https://doi.org/10.1093/asj/sjz271 Messingschlager, T. V., & Appel, M. (2025). Algorithmic bias in image-generating artificial intelligence: prevalence and user perceptions. Information, Communication & Society , 1-23. https://doi.org/10.1080/1369118X.2025.2584146 Nakao, K., & Treas, J. (1992). The 1989 socioeconomic index of occupations: Construction from the 1989 occupational prestige scores (Vol. 74). National Opinion Research Center Chicago. Noble, S., & Roberts, S. (2019). T echnological Elites, the Meritocracy, and Postracial Myths in Silicon Valley. . RACISM POSTRACE. , 6 . Parzer, V., Sjöholm, K., Brix, J. M., Svensson, P. A., Ludvik, B., & Taube, M. (2021). Development of a BMI-Assigned Stunkard Scale for the Evaluation of Body Image Perception Based on Data of the SOS Reference Study. Obes Facts , 14 (4), 397-404. https://doi.org/10.1159/000516991 Scagliusi, F. B., Alvarenga, M., Polacow, V. O., Cordás, T. A., de Oliveira Queiroz, G. K., Coelho, D.,…Lancha, A. H. (2006). Concurrent and discriminant validity of the Stunkard's figure rating scale adapted into Portuguese. Appetite , 47 (1), 77-82. https://doi.org/https://doi.org/10.1016/j.appet.2006.02.010 Silventoinen, K., Lahtinen, H., Kilpi, F., Morris, T. T., Davey Smith, G., & Martikainen, P. (2024). Socio-economic differences in body mass index: the contribution of genetic factors. International Journal of Obesity , 48 (5), 741-745. https://doi.org/10.1038/s41366-024-01459-w Stunkard, A. J., Sorensen, T., & Schulsinger, F. (1983). Use of the Danish adoption register for the study of obesity and thinness. In S. S. Kety, L. P. Rowland, R. L. Sidman, & S. W. Matthysse (Eds.), The genetics of neurological and psychiatric disorders (pp. 115-120). Raven. Szymański, P., Lipczyńska, M., & Górska, A. M. (2025). From data to perception: visualizing bias in artificial intelligence-generated images. European Heart Journal , 46 (19), 1781-1782. https://doi.org/10.1093/eurheartj/ehae850 The Oxford English dictionary. (2026). The Oxford English dictionary. In OED online . https://www.oed.com/ Thomas, J., Aljedawi, Y., AlBeyahi, F., & Rashid, M. (2025). Punctuated Equilibrium: Exploring Public Attitudes towards Artificial Intelligence across 35 Countries. In M. A. Kuhail, M. Mohamad, R. Hammad, & M. Bahja (Eds.), Unleashing User Innovation: Multidisciplinary Perspectives on End-User Development and Generative AI . Taylor & Francis Group. https://doi.org/doi.org/10.1201/9781003485698 Thompson, J. K., & Altabe, M. N. (1991). Psychometric qualities of the Figure Rating Scale. International Journal of Eating Disorders (10), 615–619. Thompson, J. K., Heinberg, L. J., Altabe, M., & Tantleff-Dunn, S. (1999). Exacting beauty: Theory, assessment, and treatment of body image disturbance [doi:10.1037/10312-000]. American Psychological Association. https://doi.org/10.1037/10312-000 Thorndike, E. L. (1920). A constant error in psychological ratings. Journal of Applied Psychology , 4 (1), 25-29. https://doi.org/10.1037/h0071663 Whang, S. E., Roh, Y., Song, H., & Lee, J.-G. (2023). Data collection and quality challenges in deep learning: a data-centric AI perspective. The VLDB Journal , 32 (4), 791-813. https://doi.org/10.1007/s00778-022-00775-9 World Health Organization. (2022). Global Health Observatory Data Repository: Prevalence of stunting, wasting, and overweight [Data set] (https://doi.org/https://www.who.int/data/gho/data/indicators Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9266968","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":614593605,"identity":"003ddaf6-0750-4b8a-8df4-a0266d92dc02","order_by":0,"name":"Justin Thomas","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0002-8118-2947","institution":"University of Liverpool","correspondingAuthor":true,"prefix":"","firstName":"Justin","middleName":"","lastName":"Thomas","suffix":""},{"id":614593606,"identity":"69ebc5f6-5d14-420f-aaa6-e5aefcbba540","order_by":1,"name":"Fahad Al-Beyahi","email":"","orcid":"","institution":"King Abdulaziz Center for World Culture (Ithra), Saudi Arabia","correspondingAuthor":false,"prefix":"","firstName":"Fahad","middleName":"","lastName":"Al-Beyahi","suffix":""},{"id":614593607,"identity":"0d6ecee2-ad48-45ce-880c-7dac90300ba0","order_by":2,"name":"Najud Al-Dugheithir","email":"","orcid":"","institution":"King Abdulaziz Center for World Culture (Ithra), Saudi Arabia","correspondingAuthor":false,"prefix":"","firstName":"Najud","middleName":"","lastName":"Al-Dugheithir","suffix":""},{"id":614593608,"identity":"cca48a84-f279-4de9-b06a-a1643a3d8066","order_by":3,"name":"Mohammad Amin Kuhail","email":"","orcid":"","institution":"Zayed University","correspondingAuthor":false,"prefix":"","firstName":"Mohammad","middleName":"Amin","lastName":"Kuhail","suffix":""}],"badges":[],"createdAt":"2026-03-30 12:38:58","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9266968/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9266968/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106259369,"identity":"8e3c6838-bf74-45fe-890e-a7e864a0218e","added_by":"auto","created_at":"2026-04-06 20:17:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":132244,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eUsed to Assign Body Size to AI-generated Depictions of Women, Including the Corresponding Estimated BMI and Weight Status Based on Parzer et al.(2021)\u003c/em\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9266968/v1/aea19eef072690eb82ccbf5d.png"},{"id":106259370,"identity":"6134677e-83e7-4d35-b35d-ea73320eea86","added_by":"auto","created_at":"2026-04-06 20:17:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":220871,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eExamples of AI-generated Depictions of Women Across Levels of Occupational Prestige.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9266968/v1/635305a0c97ab1d13f8bf52d.png"},{"id":106402826,"identity":"6363fdef-77d1-4e8f-b61b-a3ac1277bb81","added_by":"auto","created_at":"2026-04-08 09:12:58","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":86725,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eScreenshot of One of 36 Body Size Rating Trials\u003c/em\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9266968/v1/9015d82088b13cab7d25d93f.png"},{"id":106259371,"identity":"71a71501-0cea-412f-a310-d4f860e9a909","added_by":"auto","created_at":"2026-04-06 20:17:14","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":55503,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003ePercentage of AI-generated Depictions of Women Rated as Underweight, Normal Weight and Overweight Compared with WHO Population Estimates of Weight Status among US Women\u003c/em\u003e\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9266968/v1/ab2f8ec98059980a8bc224b7.png"},{"id":106259372,"identity":"ce33f8f4-e4d9-4ed2-add8-a9840b9283ff","added_by":"auto","created_at":"2026-04-06 20:17:14","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":212049,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eThe AI depictions of women that participants most frequently rated as underweight and overweight, respectively.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9266968/v1/0f2a32574c2d395b8da833b9.png"},{"id":106259373,"identity":"2a4a730e-37f6-4cbd-87d8-c203cf49ff27","added_by":"auto","created_at":"2026-04-06 20:17:14","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":87541,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eThe Percentage of Underweight, Normal Weight, and Overweight by AI Platform and by Occupational Prestige\u003c/em\u003e\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-9266968/v1/b7f6fa3396c989d7eaddddda.png"},{"id":106259374,"identity":"3da35332-2b5f-4bb0-98be-b311c6c82791","added_by":"auto","created_at":"2026-04-06 20:17:14","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":44823,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eMean Body Size Rating by Platform and by Occupational Prestige\u003c/em\u003e\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-9266968/v1/23c2338bcf38e321b720120b.png"},{"id":106406745,"identity":"bc7b1bd2-2375-421a-ad15-d9b55b77bcb7","added_by":"auto","created_at":"2026-04-08 09:33:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1363933,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9266968/v1/c38588ce-646e-4263-9a26-bdfc76cc00e0.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eUnderweight and Overrepresented: Body Size Bias in AI-Generated Depictions of Women\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe launch of OpenAI's ChatGPT in November 2022 introduced the internet-using public to generative AI. Within eight weeks of the platform’s launch, it had gained 100 million users (mostly in the USA), vastly surpassing the growth rates of all earlier online platforms, making it the fastest-growing consumer software application in history(Thomas et al., 2025). Subsequently, other companies launched similar services (Google Gemini, MetaAI assistant, Deepseek), and several began offering text-to-image features, allowing users to prompt AI to generate images from natural-language inputs.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThese text-to-image AI tools are rapidly changing how people source and produce images for various purposes, from personal enjoyment to building visual brand identities. According to market intelligence reports, around 34 million such images are generated each day, and it is estimated that since 2022, over 15 billion AI images have been created (DigitalSilk, 2026). Given the rapid proliferation of AI tools and continuous improvements in their output quality, AI-generated images are likely to become a major part of the visual environments people encounter daily (Messingschlager \u0026amp; Appel, 2025).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOne of the challenges of such a scenario is that AI can introduce or reinforce existing social biases at scale. Such algorithmic bias, like bias in general, typically involves an inclination toward preferencing one individual or group over another (The \u0026nbsp; Oxford \u0026nbsp;English dictionary, 2026) or, in the more abstract sense the “..deviation of a measured or calculated quantity from its actual (true) value, such that the measurement or calculation is unrepresentative of the item of interest(\u003cem\u003eAPA Dictionary of Psychology\u003c/em\u003e, 2007). For example, if the actual (true) rate of female chief executive officers (CEOs) was 12% and our image-to-text algorithm, when repeatedly prompted to “draw a CEO”, generated female CEOs at a rate of 0.12%, we could describe this as representational gender bias. The origins of such biases within AI algorithms are usually linked to unrepresentative training data, meaning the data used to train the system fails to adequately reflect the diversity of the real-world population it aims to serve. Furthermore, the designers of such systems may inadvertently embed their own biases and assumptions within their creations(Noble \u0026amp; Roberts, 2019).\u003c/p\u003e\n\u003cp\u003eResearchers exploring biases in AI text-to-image systems have begun documenting cases. In one study, nine different text-to-image platforms were used to visually represent professionals working in the prestigious fields of law, medicine, engineering, and scientific research(Gorska \u0026amp; Jemielniak, 2023). Based on viewer categorisation of the resultant output, 76% of the images were deemed male, while only 8% were judged female. The bias persisted across all professions, but the extent of the gender skew varied greatly by AI platforms.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAnother study focused specifically on the representation of occupations in science, technology, engineering, and maths (STEM) fields. This study asked Midjourney (an AI service specialised in image generation) to “Imagine a portrait of a [insert scientist/person in science],” where the inserted word was drawn from a list of STEM occupations, e.g., biologist, physicist. Of the 84 images generated, all were of white males, suggesting a particularly extreme representational bias against women and people of colour(Messingschlager \u0026amp; Appel, 2025). Even when the AI was explicitly prompted to depict female scientists, these were typically portrayed as much younger than their male counterparts. In this context, youth might be interpreted as “less experienced”.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe bias towards portraying females as younger than their male counterparts was also observed in a study of 1.4 million digital images(Guilbeault et al., 2025). Despite no systematic age differences between women and men in the workforce according to the US Census, the study found that women were portrayed as younger than men across a wide range of occupations, with the bias being even more evident in higher-status roles. Another study examined the representation of gender and ethnic markers in AI-generated imagery across occupations using three text-to-image platforms, discovering that both ethnicity and gender were over- or underrepresented to varying extents, thereby reinforcing existing social stereotypes(Luccioni et al., 2023). For instance, all three platforms typically depicted dentists as male (83.57% - 62.50%), whereas dental assistants were almost exclusively portrayed as females (96.79% - 94.34%). A similar trend was observed for doctors (98.09%–87.14% male) and nurses (100%–98.06% female).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eExploring the potential bias- and stereotype-amplifying implications of text-to-image AI in the context of global health imagery, one study sought to craft prompts that would challenge prevalent global health tropes/stereotypes, such as the so-called “white saviour” trope(Alenichev et al., 2023). To this end, Midjourney was tasked with depicting images of Black African doctors attending to sick White children. Despite the use of prompts crafted to challenge prevalent stereotypes, the AI invariably, across 300 images, rendered the child patients as Black, even though the prompts explicitly specified White children. Similarly, Black African doctors were sometimes depicted as White, despite that being the stereotype the researchers had sought to invert (Alenichev et al., 2023). The Authors concluded that “..despite the supposed enormous generative power of AI, it proved incapable of avoiding the perpetuation of existing inequality and prejudice.” (p. 1496).\u003c/p\u003e\n\u003cp\u003eAlmost invariably, the existing literature on bias in AI imagery warns of its potential to perpetuate stereotypes and further entrench social biases, inequalities, and power imbalances (Alenichev et al., 2023; Guilbeault et al., 2025; Szymański et al., 2025). Whether intentional or inadvertent, biases in AI-generated imagery can impact public perceptions, shape opinions, and affect the assumptions individuals make about others or themselves. To date, however, much of this research has remained focused on gender and ethnicity-related biases. We are unable to find any studies that have systematically examined how AI’s text-to-image functions represent body shape and size, which can also serve as a rough approximation of weight status(Parzer et al., 2021). This is an important focus given the extensive and widely reported links between media imagery, body image dissatisfaction and eating disorders, especially (although not exclusively) among females. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSince the 1990s, eating disorders have become a significant and increasingly global public health problem \u0026nbsp;(Austin, 2012; Gordon, 2000). Those experiencing such disorders demonstrate heightened mortality rates, with the risk of death more than three times higher than that of the general population, while among those diagnosed with anorexia nervosa, the mortality rate rises to fivefold \u0026nbsp;(Krug et al., 2025).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe onset and maintenance of eating disorders such as anorexia and bulimia nervosa is multifaceted, involving biological, psychological and socio-cultural factors(American Psychiatric Association, 2022). Within the context of sociocultural influences, however, extensive research indicates that the promotion of increasingly thin and unattainable female body image ideals in the media can play a significant role in the onset of such disorders(Gordon, 1990, 2001; Groesz et al., 2002). The proposed mechanism of action involves self-comparison with media-promoted ideals, which fosters body image dissatisfaction (Laker \u0026amp; Waller, 2022) and maladaptive body image attitudes, such as overvaluing thinness(Fairburn et al., 2003). In turn, body image dissatisfaction and maladaptive body image attitudes can give rise to maladaptive weight-control measures (Larson et al., 2021) and eating disorders such as anorexia and bulimia nervosa(American Psychiatric Association, 2013).\u003c/p\u003e\n\u003cp\u003eThe idea of bias toward portraying thinner female body image ideals in traditional media is well-documented. A seminal study by Garner et al. (1980) reported decreases in the weight and body dimensions of US beauty pageant contestants between 1959 and 1978, coinciding with rising weight norms for US women. More recently, Maymone et al (2019) reported the same patterns for Victoria’s Secret models between 1995 and 2018. \u0026nbsp;This bias toward celebrating thinner female body images, typically depicted in the media, highlights discrepancies between society-wide body image ideals and anthropometric realities(Thompson et al., 1999). The current rate of overweight and obesity among US women stands at 70.3%\u0026nbsp;[67.1 – 73.3], according to\u0026nbsp;the\u0026nbsp;World Health Organisation(2022), with similar rates reported by the US National Centre for Health Statistics(Fryar et al., 2026).\u0026nbsp;Discrepancies between body image aspirations/ideals and anthropometric realities or perceptions are central to most definitions of body image dissatisfaction(Cash \u0026amp; Szymanski, 1995; Garner, 1997; Khalaf et al., 2011; Thompson \u0026amp; Altabe, 1991; Thompson et al., 1999).\u003c/p\u003e\n\u003cp\u003eAI’s text-to-image function threatens to further amplify the existing body image biases, further exacerbating societal levels of body image dissatisfaction and inadvertently promoting eating disorders. In this study, we aim to explore the extent to which text-to-image AI depicts women as either underweight, normal weight, or overweight. \u0026nbsp;Given that AI systems are typically trained on images and captions scraped from the public internet, including digitised legacy media(Whang et al., 2023), we make the following hypotheses.\u003c/p\u003e\n\u003cp\u003eHypothesis 1: The rate of perceiving AI-depicted women as underweight will be higher than the estimated age-adjusted prevalence of underweight for women in the US population. Hypothesis 2: The rate of perceiving AI-depicted women as underweight will be higher than the estimated age-adjusted prevalence of underweight for women in the US population. Furthermore, considering previously observed platform-level differences in gender bias, we also propose the following hypothesis. \u0026nbsp;Hypothesis 3: Perceived body size of AI-depicted women will vary by platform. Finally, previous studies have documented increased levels of gender (male) and age (young female) bias linked with higher occupational status. The same might be true for thinness. We suspect an algorithmic variant of the “halo effect” (Thorndike, 1920) is at work, whereby socially desirable traits such as thinness, youthfulness, and occupational prestige are combined for females. Based on these ideas, we formulate the following hypothesis. Hypothesis 4: AI-depicted women in roles with higher occupational prestige will be rated as having thinner body sizes.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eParticipants\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eParticipants were recruited through Prolific, a crowdsourcing platform dedicated to social science research. Prolific is among the most widely used recruitment platforms for research in this field and has been reported to deliver higher-quality data than similar services (Peer et al., 2017; Stritch et al., 2025). The present non-probability quota sample consisted of 129 young women (aged 18–34) living in the United States. The focus on young women was motivated by previous research showing that this age group and gender are particularly vulnerable to body dissatisfaction and engaging in body-focused social comparison online, that is, comparing one's physical appearance with digital images of others (Merino et al., 2024). The focus on US residents was justified by the relatively widespread use of AI in US society and by the fact that the AI platforms under examination are founded and hosted in the US. Participants were selected from a pool of 22,689 eligible individuals and were allowed to participate only on a personal computer or tablet. In addition to age, they were asked to report their height, weight, and current employment status. Most (73.43%) were employed, while 10.15% were engaged in education, 14.06% were unemployed, and 2.34% identified their employment status as other. Table 1 provides details of all other participant characteristics.\u003c/p\u003e\n\u003cp\u003eTable 1 \u003cem\u003eSample Characteristics\u003c/em\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e26.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e26.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e3.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e7.89\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMedian\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e26.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e26.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eIQR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e9.65\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e16.73\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMax\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e58.58\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eMeasures and Materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStunkard Figure Rating Scale (FRS)\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe original FRS (Stunkard et al., 1983) consists of nine silhouette drawings of body shapes on a gradient from underweight to obese, with separate figure sets for males and females. Previous studies report good convergent and discriminant validity for the scale, as well as test-retest reliability, across diverse populations (Lo et al., 2011; Scagliusi et al., 2006). Each figure is assigned a number/score, 1 to 9; higher scores indicate larger body sizes. \u0026nbsp;Simple to administer, this measure has been widely used to assess one's current body size, one’s body size aspirations, and to assign perceived body size or weight status to others. For example, Cardinal et al. (2006) used the FRS to assign figure ratings to adult women viewed on videotape. In this instance, observer ratings were correlated with the actual BMIs of the women being rated, r [70] = .91. Similarly, in the present study, rather than using the scale for own-body evaluation, participants were asked to rate the body size of AI-generated depictions of women. The current study used an abbreviated six-silhouette version of the FRS, which facilitated a more parsimonious categorisation of the AI-generated depictions as either underweight, normal weight, or overweight. Based on Parzer et al. (2021), a large population study exploring the correspondence between FRS silhouettes and clinically assessed BMI, silhouette 1 was taken to represent underweight, silhouettes 2, 3, and 4 represented normal weight, and silhouettes 5 and 6 were categorised as overweight. Figure 1 details the estimated correspondence between the silhouettes, BMI, and weight status.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAI-Generated Depictions of Women\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe chose three general-purpose AI platforms/assistants for image generation based on their current popularity, free access and their provision of text-to-image features. These platforms were ChatGPT 5.3, Google AI Studio and Meta AI assistant. Including additional platforms was beyond the scope of this initial investigation.\u003c/p\u003e\n\u003cp\u003eTo generate images, we used a standardised natural language prompt across all platforms: “Draw a female (insert occupation – e.g., firefighter), head to toe, facing the viewer”. In the few cases when the AI hallucinated, for example, depicting a woman with three legs, it was instructed to “try again”. There were only three instances where this additional step was required. All images were generated between 1\u003csup\u003est\u003c/sup\u003e and 14\u003csup\u003eth\u003c/sup\u003e September 2025.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOccupational Prestige\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe AI-generated images spanned three categories: high, middle and low occupational prestige. This categorisation was based on data from the National Opinion Research Centre (NORC), in which occupational categories listed in the 1980 US Census were rated on a 0-100 prestige scale by approximately 1,200 independent raters (Hauser \u0026amp; Warren, 1997; Nakao \u0026amp; Treas, 1992). This list contains 740 occupations, and we randomly selected 12 occupations from the upper (Lawyer, CEO), middle (Librarian, Chef), and lower (e.g., bus driver, seamstress) terciles of the occupational prestige list. \u0026nbsp;We excluded any occupations in which body shape and size can be associated with the role, for example, athletes and cabin crew. There were 36 images in total, equally distributed across the occupational prestige categories and AI platforms (3 x 12). Figure 2 provides examples of images used in the study from each occupational prestige category.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eProcedure\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study protocol was conducted online and received prospective ethics approval from the institution's internal review board. Participants received a brief overview of the study and were informed that the data collection would remain anonymous. They then could indicate their consent to participate by clicking 'continue'. After answering demographic questions, participants completed the body size rating task, which involved 36 trials in which they used the FRS to estimate an approximate body size for each AI-generated image (see figure 3).\u003c/p\u003e\n\u003cp\u003eThe presentation order of AI-generated images was arranged using a permutation that prevented consecutive repetition of occupational prestige levels or image sources (i.e., AI platforms). All data were collected on 9 November 2025. The average task completion time was 7 minutes 43 seconds, with all participants completing every trial.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eIn this study, 16.64% of the AI-generated depictions of women were rated as overweight (body image figures 5-6), while 7.76% were perceived as being underweight (Silhouette 1). By comparison, based on age-standardised estimates from the World Health Organisation(2022), 70.3% [67.1 – 73.3] of US females are overweight (BMI \u0026gt;= 25), while the age-standardised rate of underweight (BMI \u0026lt; 18) among US females is 2.3% [1.5 – 3.3]. Figure 4 below visually compares the percentages of AI-generated female images rated as underweight, normal weight, and overweight with US weight-status norms reported by the World Health Organisation (WHO) in 2022. \u0026nbsp;These findings lend support for hypotheses 1 and 2. Suggesting that, in AI portrayals of women, underweight is perceived more frequently than in the general population, and conversely, overweight is perceived less frequently than in the general population.\u003c/p\u003e\n\u003cp\u003eThe image most frequently classified as underweight, that is, the image that most viewers rated as matching the FRS silhouette 1, was image 7, see figure 5 below. Previous research suggests that this body-silhouette corresponds to a BMI of 17.88 (Parzer et al., 2021). By contrast, the image most frequently classified as underweight, that is, the image that most viewers rated as matching the FRS silhouette 5 or 6, was image 24, see figure 5 below. Previous research suggests that this body silhouette corresponds to a BMI of 25.68–28.75 (Parzer et al., 2021).\u003c/p\u003e\n\u003cp\u003eAI body-size ratings were uncorrelated with participants' self-reported BMI; however, younger participants tended to provide higher (larger) body-size ratings, \u003cem\u003er\u0026nbsp;\u003c/em\u003e[127] = .244, \u003cem\u003ep\u003c/em\u003e = .005. Supportive of hypothesis 3, the body size ratings for the AI-generated depictions of women varied widely across platforms, with the frequency of underweight ratings lowest on ChatGPT, which also received the highest frequency of overweight ratings for its generated images. The percentage of underweight, normal weight, and overweight by platform and occupational prestige is detailed in Figure 6\u003c/p\u003e\n\u003cp\u003eThe mean figure rating score for the ChatGPT-generated images was 4.10 (\u003cem\u003eSD\u0026nbsp;\u003c/em\u003e= 1.19), for GoogleAI it was 2.65 (\u003cem\u003eSD\u003c/em\u003e = 0.95) and for MetaAI it was 2.78 (\u003cem\u003eSD\u003c/em\u003e = 1.15). A repeated-measures ANOVA was performed to evaluate the effect of the AI platform on body size ratings. Mauchly’s test indicated that the assumption of sphericity was violated, χ²(2) = 125.59, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001. Degrees of freedom were corrected using Greenhouse-Geisser estimates (ε = .92). The effect of the AI platform on body size ratings was significant, \u003cem\u003eF\u003c/em\u003e(2, 1535) = 915.98, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001, partial η² = .37. Post hoc pairwise comparisons indicated that GoogleAI images were perceived as thinner than those generated by MetaAI (\u003cem\u003ep\u003c/em\u003e \u0026lt; .001) and ChatGPT (\u003cem\u003ep\u003c/em\u003e \u0026lt; .001). Supporting hypotheses 3 and 4, Figure 7 plots the means of body size ratings across the three AI platforms and occupational prestige.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe mean figure rating score for the AI-generated images of women working in high-status occupations was 2.66 (\u003cem\u003eSD\u003c/em\u003e = 1.15); for middle-status occupations, 3.12 (SD = 1.16); and for low-status occupations, 3.758 (\u003cem\u003eSD\u003c/em\u003e = 1.29). A repeated-measures ANOVA was performed to evaluate the effect of occupational status on perceptions of body size. Mauchly’s test indicated that the assumption of sphericity was violated, χ²(2) = 95.48, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001. Degrees of freedom were corrected using Greenhouse-Geisser estimates (ε = .94). The effect of the AI platform on body size ratings was significant, \u003cem\u003eF\u003c/em\u003e(2, 1535) = 325.98, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001, partial η² = .17. Post hoc pairwise comparisons indicated that images of women in high status occupations were perceived as thinner than those in middle status occupations (\u003cem\u003ep\u003c/em\u003e \u0026lt; .001) and low status occupations (\u003cem\u003ep\u003c/em\u003e \u0026lt; .001).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study examined the perceived body size and, by extension, perceived weight status of AI-generated images of women. As hypothesised, the proportion of perceived underweight was much higher (more than three times greater) than that reported in the general adult female population of the USA. Similarly, the proportion of AI-generated images of women perceived as overweight was substantially lower than US population estimates. This pattern was even more pronounced for depictions of women working in more prestigious occupations. A quantitative analysis of the figure ratings for the AI images showed that portrayals of women in high-prestige roles (e.g., CEO, Lawyer, Architect) were rated as significantly thinner than those in roles associated with middle and lower levels of occupational prestige (e.g., Teacher, Garbage Collector). The links with occupation prestige may reflect the tendency for those of lower socioeconomic status to have higher rates of overweight and obesity in Western societies (Silventoinen et al., 2024). However, it might also reflect a clustering of socially desirable attributes (youth, beauty, occupational prestige, and thinness) akin to the halo effect(Thorndike, 1920), in which the perception of one positive (socially desirable) trait leads to unfounded inferences about the presence of other positive traits. \u0026nbsp;At least one study suggests AI can be prone to such halo effects (Gulati et al., 2025), and the idea is certainly worthy of future research.\u003c/p\u003e\n\u003cp\u003eOverall, these findings support the hypothesis that text-to-image AI systematically overrepresents very thin (potentially underweight) women while underrepresenting overweight women, at least in comparison to current US weight norms. This body-size-related representational skew extends the list of biases (gender, age, and ethnicity) identified in earlier studies of AI’s text-to-image function (Alenichev et al., 2023; Guilbeault et al., 2025; Szymański et al., 2025). The fact that the three platforms differed significantly in the extent to which they produced body-size bias suggests that engineering solutions could minimise or even eliminate such bias. These platform-level differences have also been observed in relation to the magnitude of gender-biased portrayals (Gorska \u0026amp; Jemielniak, 2023), again suggesting that such biases are, to some extent, platform-specific and not beyond attenuation. Attempts at bias correction, however, raise questions about what the optimum outputs should be: a true representation of reality, engineered equality, or overrepresentation of historically marginalised groups? From a public health perspective, we argue that, in the context of body size, the optimal default would be a body size that reflects a healthy weight status, avoiding the extremes of underweight and overweight. If such extreme images are desired, then users can add additional prompts. Currently, at least two of the AI platforms are generating images of women that are widely perceived as underweight, at rates far exceeding population-level estimates of this weight status. \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIt appears likely that Western society’s promotion of thinner, more tubular female body ideals since the 1960s (Bell, 1985; Bruch, 1974; Gordon, 2000)\u0026nbsp;is well reflected in the legacy data used to train current AI systems. However, explicit design or aesthetic choices made by those developing such systems might also contribute to overrepresenting women broadly perceived as underweight. For instance, analysis of female video game characters indicates that technology developers often intentionally portray these characters with ultra-thin, unrealistic body types (De la Torre-Sierra \u0026amp; Guichot-Reina, 2025). Regardless of how this body-size bias has been established in text-to-image AI systems, it seems likely that, if left unaddressed, it will perpetuate and possibly even amplify unrealistic and unhealthy body image ideals. As the use of AI text-to-image systems grows exponentially, we may inadvertently promote body image dissatisfaction, upward social comparison(Laker \u0026amp; Waller, 2022), and the “overvaluing of thinness”(Fairburn et al., 2003), all of which appear to increase the risk of eating disorders among vulnerable individuals(American Psychiatric Association, 2022). If left unchecked, we risk inadvertently promoting anorexigenic AI.\u003c/p\u003e\n\u003cp\u003eThis is an important and emerging concern that has, to date, received limited research attention. This study provides a preliminary examination of the default female body sizes generated by AI text-to-image functionality on three widely used platforms. The study has several important limitations. Firstly, the exploratory nature of this study precluded the inclusion of more platforms, more images, and a larger sample size. Although adequate for the current study, future explorations of this question would benefit from larger samples and more images drawn from a broader array of platforms. Similarly, the current study’s exclusive focus on women limits generalizability and the ability to contrast and compare with portrayals of male body size. Additionally, the use of a crude silhouette figure rating scale (FRS) limits measurement sensitivity and accuracy. For instance, the FRS has only one silhouette corresponding to underweight, whereas there are multiple silhouettes for overweight and normal weight. Future explorations of body-size bias in AI imagery should extend to males, and more sophisticated, sensitive measures should be used in the image-rating process.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSetting limitations aside, the present study reveals body-size bias in AI-generated portrayals of women. These insights can inform health policies targeting the digital determinants of health, support the efforts of eating disorder advocacy groups, and assist those in the technology sector in monitoring and promoting ethical and safe AI practices.\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAlenichev, A., Kingori, P., \u0026amp; Grietens, K. P. (2023). Reflections before the storm: the AI reproduction of biased imagery in global health visuals. \u003cem\u003eThe Lancet Global Health\u003c/em\u003e,\u003cem\u003e 11\u003c/em\u003e(10), e1496-e1498. https://doi.org/10.1016/S2214-109X(23)00329-7\u003c/li\u003e\n\u003cli\u003eAmerican Psychiatric Association. (2013). \u003cem\u003eDiagnostic and statistical manual of mental disorders : DSM-5\u003c/em\u003e. American Psychiatric Association.\u003c/li\u003e\n\u003cli\u003eAmerican Psychiatric Association. (2022). \u003cem\u003eDiagnostic and statistical manual of mental disorders : DSM-5-TR\u003c/em\u003e. American Psychiatric Association.\u003c/li\u003e\n\u003cli\u003e\u003cem\u003eAPA Dictionary of Psychology\u003c/em\u003e. (2007). American Psychological Association.\u003c/li\u003e\n\u003cli\u003eAustin, S. B. (2012). A public health approach to eating disorders prevention: It\u0026rsquo;s time for public health professionals to take a seat at the table. \u003cem\u003eBMC Public Health\u003c/em\u003e,\u003cem\u003e 12\u003c/em\u003e(1), 854. https://doi.org/10.1186/1471-2458-12-854\u003c/li\u003e\n\u003cli\u003eBell, R. M. (1985). \u003cem\u003eHoly Anorexia\u003c/em\u003e. University of Chicago Press.\u003c/li\u003e\n\u003cli\u003eBruch, H. (1974). \u003cem\u003eEating Disorders\u003c/em\u003e. Routledge \u0026amp; Kegan Paul.\u003c/li\u003e\n\u003cli\u003eCardinal, T. M., Kaciroti, N., \u0026amp; Lumeng, J. C. (2006). The figure rating scale as an index of weight status of women on videotape. \u003cem\u003eObesity (Silver Spring)\u003c/em\u003e,\u003cem\u003e 14\u003c/em\u003e(12), 2132-2135. https://doi.org/10.1038/oby.2006.249\u003c/li\u003e\n\u003cli\u003eCash, T. F., \u0026amp; Szymanski, M. L. (1995). The Development and Validation of the Body-Image Ideals Questionnaire. \u003cem\u003eJournal of Personality Assessment\u003c/em\u003e,\u003cem\u003e 64\u003c/em\u003e(3). https://doi.org/10.1207/s15327752jpa6403_6\u003c/li\u003e\n\u003cli\u003eDe la Torre-Sierra, A. M., \u0026amp; Guichot-Reina, V. (2025). Women in Video Games: An Analysis of the Biased Representation of Female Characters in Current Video Games. \u003cem\u003eSexuality \u0026amp; Culture\u003c/em\u003e,\u003cem\u003e 29\u003c/em\u003e(2), 532-560. https://doi.org/10.1007/s12119-024-10286-0\u003c/li\u003e\n\u003cli\u003eFairburn, C. G., Cooper, Z., \u0026amp; Shafran, R. (2003). Cognitive behaviour therapy for eating disorders: a \u0026quot;transdiagnostic\u0026quot; theory and treatment. \u003cem\u003eBehav Res Ther\u003c/em\u003e,\u003cem\u003e 41\u003c/em\u003e(5), 509-528. https://doi.org/10.1016/s0005-7967(02)00088-8\u003c/li\u003e\n\u003cli\u003eFryar, C. D., Afful, J., \u0026amp; Saif, N. (2026). \u003cem\u003eHealth E-Stat 111: Prevalence of Overweight, Obesity, and Severe Obesity Among Adults Age 20 and Older: United States, 1960\u0026ndash;1962 Through August 2021\u0026ndash;August 2023\u003c/em\u003e (National Health and Nutrition Examination Survey, Issue. https://stacks.cdc.gov/view/cdc/174643\u003c/li\u003e\n\u003cli\u003eGarner, D. M. (1997). Body image: an in-depth look at how we see ourselves. \u003cem\u003ePsychology Today\u003c/em\u003e, (01 Febuary 1997).\u003c/li\u003e\n\u003cli\u003eGordon, R. A. (1990). \u003cem\u003eAnorexia and Bulimia\u003c/em\u003e. Blackwell.\u003c/li\u003e\n\u003cli\u003eGordon, R. A. (2000). \u003cem\u003eEating Disorders: Anatomy of a Social Epidemic\u003c/em\u003e (2nd ed.). Blackwell.\u003c/li\u003e\n\u003cli\u003eGordon, R. A. (2001). Eating disorders East and West: A culture-bound syndrome unbound. In M. Nasser, M. A. Katzman, \u0026amp; R. A. Gordon (Eds.), \u003cem\u003eEating Disorders and Cultures in Transition\u003c/em\u003e. Brunner-Routledge.\u003c/li\u003e\n\u003cli\u003eGorska, A. M., \u0026amp; Jemielniak, D. (2023). The invisible women: uncovering gender bias in AI-generated images of professionals. \u003cem\u003eFeminist Media Studies\u003c/em\u003e,\u003cem\u003e 23\u003c/em\u003e(8), 4370-4375. https://doi.org/10.1080/14680777.2023.2263659\u003c/li\u003e\n\u003cli\u003eGroesz, L. M., Levine, M. P., \u0026amp; Murnen, S. K. (2002). The effect of experimental presentation of thin media images on body satisfaction: a meta-analytic review. \u003cem\u003eInt J Eat Disord\u003c/em\u003e,\u003cem\u003e 31\u003c/em\u003e(1), 1-16. https://doi.org/10.1002/eat.10005\u003c/li\u003e\n\u003cli\u003eGuilbeault, D., Delecourt, S., \u0026amp; Desikan, B. S. (2025). Age and gender distortion in online media and large language models. \u003cem\u003eNature\u003c/em\u003e,\u003cem\u003e 646\u003c/em\u003e(8087), 1129-1137. https://doi.org/10.1038/s41586-025-09581-z\u003c/li\u003e\n\u003cli\u003eGulati, A., D\u0026apos;Inc\u0026agrave;, M., Sebe, N., Lepri, B., \u0026amp; Oliver, N. (2025). \u003cem\u003eBeauty and the Bias: Exploring the Impact of Attractiveness on Multimodal Large Language Models\u003c/em\u003e. https://doi.org/10.48550/arXiv.2504.16104\u003c/li\u003e\n\u003cli\u003eHauser, R. M., \u0026amp; Warren, J. R. (1997). 4. Socioeconomic Indexes for Occupations: A Review, Update, and Critique. \u003cem\u003eSociological Methodology\u003c/em\u003e,\u003cem\u003e 27\u003c/em\u003e(1), 177-298. https://doi.org/10.1111/1467-9531.271028\u003c/li\u003e\n\u003cli\u003eKhalaf, A., Westergren, A., Berggren, V., \u0026amp; Al-Hazzaa, H. (2011). PP201-SUN Discrepancies between the actual, perceived and ideal body image among female university students in south western Saudi Arabia. \u003cem\u003eClinical Nutrition Supplements\u003c/em\u003e,\u003cem\u003e 6\u003c/em\u003e, 99-100.\u003c/li\u003e\n\u003cli\u003eKrug, I., Liu, S., Portingale, J., Croce, S., Dar, B., Obleada, K.,\u0026hellip;Fuller-Tyszkiewicz, M. (2025). A meta-analysis of mortality rates in eating disorders: An update of the literature from 2010 to 2024. \u003cem\u003eClinical Psychology Review\u003c/em\u003e,\u003cem\u003e 116\u003c/em\u003e, 102547. https://doi.org/https://doi.org/10.1016/j.cpr.2025.102547\u003c/li\u003e\n\u003cli\u003eLaker, V., \u0026amp; Waller, G. (2022). Does comparison of self with others influence body image among adult women? An experimental study in naturalistic settings. \u003cem\u003eEat Weight Disord\u003c/em\u003e,\u003cem\u003e 27\u003c/em\u003e(2), 597-604. https://doi.org/10.1007/s40519-021-01196-3\u003c/li\u003e\n\u003cli\u003eLarson, N., Loth, K. A., Eisenberg, M. E., Hazzard, V. M., \u0026amp; Neumark-Sztainer, D. (2021). Body dissatisfaction and disordered eating are prevalent problems among U.S. young people from diverse socioeconomic backgrounds: Findings from the EAT 2010-2018 study. \u003cem\u003eEat Behav\u003c/em\u003e,\u003cem\u003e 42\u003c/em\u003e, 101535. https://doi.org/10.1016/j.eatbeh.2021.101535\u003c/li\u003e\n\u003cli\u003eLo, W. S., Ho, S. Y., Wong, B. Y., Mak, K. K., \u0026amp; Lam, T. H. (2011). Validity and test-retest reliability in assessing current body size with figure drawings in Chinese adolescents. \u003cem\u003eInt J Pediatr Obes\u003c/em\u003e,\u003cem\u003e 6\u003c/em\u003e(2-2), e107-113. https://doi.org/10.3109/17477166.2011.575146\u003c/li\u003e\n\u003cli\u003eLuccioni, A. S., Akiki, C., Mitchell, M., \u0026amp; Jernite, Y. (2023). \u003cem\u003eStable bias: evaluating societal representations in diffusion models\u003c/em\u003e Proceedings of the 37th International Conference on Neural Information Processing Systems, New Orleans, LA, USA.\u003c/li\u003e\n\u003cli\u003eMaymone, M. B., Laughter, M., Anderson, J. B., Secemsky, E. A., \u0026amp; Vashi, N. A. (2019). Unattainable Standards of Beauty: Temporal Trends of Victoria\u0026rsquo;s Secret Models from 1995 to 2018. \u003cem\u003eAesthetic Surgery Journal\u003c/em\u003e,\u003cem\u003e 40\u003c/em\u003e(2), NP72-NP76. https://doi.org/10.1093/asj/sjz271\u003c/li\u003e\n\u003cli\u003eMessingschlager, T. V., \u0026amp; Appel, M. (2025). Algorithmic bias in image-generating artificial intelligence: prevalence and user perceptions. \u003cem\u003eInformation, Communication \u0026amp; Society\u003c/em\u003e, 1-23. https://doi.org/10.1080/1369118X.2025.2584146\u003c/li\u003e\n\u003cli\u003eNakao, K., \u0026amp; Treas, J. (1992). \u003cem\u003eThe 1989 socioeconomic index of occupations: Construction from the 1989 occupational prestige scores\u003c/em\u003e (Vol. 74). National Opinion Research Center Chicago.\u003c/li\u003e\n\u003cli\u003eNoble, S., \u0026amp; Roberts, S. (2019). T echnological Elites, the Meritocracy, and Postracial Myths in Silicon Valley. . \u003cem\u003eRACISM POSTRACE.\u003c/em\u003e,\u003cem\u003e 6\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eParzer, V., Sj\u0026ouml;holm, K., Brix, J. M., Svensson, P. A., Ludvik, B., \u0026amp; Taube, M. (2021). Development of a BMI-Assigned Stunkard Scale for the Evaluation of Body Image Perception Based on Data of the SOS Reference Study. \u003cem\u003eObes Facts\u003c/em\u003e,\u003cem\u003e 14\u003c/em\u003e(4), 397-404. https://doi.org/10.1159/000516991\u003c/li\u003e\n\u003cli\u003eScagliusi, F. B., Alvarenga, M., Polacow, V. O., Cord\u0026aacute;s, T. A., de Oliveira Queiroz, G. K., Coelho, D.,\u0026hellip;Lancha, A. H. (2006). Concurrent and discriminant validity of the Stunkard\u0026apos;s figure rating scale adapted into Portuguese. \u003cem\u003eAppetite\u003c/em\u003e,\u003cem\u003e 47\u003c/em\u003e(1), 77-82. https://doi.org/https://doi.org/10.1016/j.appet.2006.02.010\u003c/li\u003e\n\u003cli\u003eSilventoinen, K., Lahtinen, H., Kilpi, F., Morris, T. T., Davey Smith, G., \u0026amp; Martikainen, P. (2024). Socio-economic differences in body mass index: the contribution of genetic factors. \u003cem\u003eInternational Journal of Obesity\u003c/em\u003e,\u003cem\u003e 48\u003c/em\u003e(5), 741-745. https://doi.org/10.1038/s41366-024-01459-w\u003c/li\u003e\n\u003cli\u003eStunkard, A. J., Sorensen, T., \u0026amp; Schulsinger, F. (1983). Use of the Danish adoption register for the study of obesity and thinness. In S. S. Kety, L. P. Rowland, R. L. Sidman, \u0026amp; S. W. Matthysse (Eds.), \u003cem\u003eThe genetics of neurological and psychiatric disorders\u003c/em\u003e (pp. 115-120). Raven.\u003c/li\u003e\n\u003cli\u003eSzymański, P., Lipczyńska, M., \u0026amp; G\u0026oacute;rska, A. M. (2025). From data to perception: visualizing bias in artificial intelligence-generated images. \u003cem\u003eEuropean Heart Journal\u003c/em\u003e,\u003cem\u003e 46\u003c/em\u003e(19), 1781-1782. https://doi.org/10.1093/eurheartj/ehae850\u003c/li\u003e\n\u003cli\u003eThe Oxford English dictionary. (2026). The Oxford English dictionary. In \u003cem\u003eOED online\u003c/em\u003e. https://www.oed.com/\u003c/li\u003e\n\u003cli\u003eThomas, J., Aljedawi, Y., AlBeyahi, F., \u0026amp; Rashid, M. (2025). Punctuated Equilibrium: Exploring Public Attitudes towards Artificial Intelligence across 35 Countries. In M. A. Kuhail, M. Mohamad, R. Hammad, \u0026amp; M. Bahja (Eds.), \u003cem\u003eUnleashing User Innovation: Multidisciplinary Perspectives on End-User Development and Generative AI\u003c/em\u003e. Taylor \u0026amp; Francis Group. https://doi.org/doi.org/10.1201/9781003485698\u003c/li\u003e\n\u003cli\u003eThompson, J. K., \u0026amp; Altabe, M. N. (1991). Psychometric qualities of the Figure Rating Scale. \u003cem\u003eInternational Journal of Eating Disorders\u003c/em\u003e(10), 615\u0026ndash;619.\u003c/li\u003e\n\u003cli\u003eThompson, J. K., Heinberg, L. J., Altabe, M., \u0026amp; Tantleff-Dunn, S. (1999). \u003cem\u003eExacting beauty: Theory, assessment, and treatment of body image disturbance\u003c/em\u003e [doi:10.1037/10312-000]. American Psychological Association. https://doi.org/10.1037/10312-000\u003c/li\u003e\n\u003cli\u003eThorndike, E. L. (1920). A constant error in psychological ratings. \u003cem\u003eJournal of Applied Psychology\u003c/em\u003e,\u003cem\u003e 4\u003c/em\u003e(1), 25-29. https://doi.org/10.1037/h0071663\u003c/li\u003e\n\u003cli\u003eWhang, S. E., Roh, Y., Song, H., \u0026amp; Lee, J.-G. (2023). Data collection and quality challenges in deep learning: a data-centric AI perspective. \u003cem\u003eThe VLDB Journal\u003c/em\u003e,\u003cem\u003e 32\u003c/em\u003e(4), 791-813. https://doi.org/10.1007/s00778-022-00775-9\u003c/li\u003e\n\u003cli\u003eWorld Health Organization. (2022). \u003cem\u003eGlobal Health Observatory Data Repository: Prevalence of stunting, wasting, and overweight [Data set]\u003c/em\u003e (https://doi.org/https://www.who.int/data/gho/data/indicators\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"artificial intelligence, body image, bias, eating disorder, online, media","lastPublishedDoi":"10.21203/rs.3.rs-9266968/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9266968/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe bias towards portraying thinner women in traditional media has been well-documented. However, there is limited research on such bias in images generated by artificial intelligence (AI). This study examined viewers' perceptions of AI-generated depictions of women, assessing whether these images were perceived as underweight, normal weight, or overweight. Using standardised prompts that did not specify body shape or size, images of women across various occupations (e.g., “draw a female lawyer”) were created using three popular generative AI tools. Participants (\u003cem\u003eN\u003c/em\u003e=129), adult women from the US, rated the body size of the AI-generated images using a shortened version of the Stunkard figure rating scale. Overall, the proportion of images perceived as underweight was more than threefold the expected rate of underweight among women in the US population. Across all platforms, images of women with greater occupational prestige were more likely to be rated as underweight. The findings suggest a systematic body size bias in text-to-image AI with implications for population-level body image dissatisfaction and eating disorders.\u003c/p\u003e","manuscriptTitle":"Underweight and Overrepresented: Body Size Bias in AI-Generated Depictions of Women","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-06 20:17:00","doi":"10.21203/rs.3.rs-9266968/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"11d6c192-58df-46c5-84b0-53e6e94a258a","owner":[],"postedDate":"April 6th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":65397268,"name":"Psychology"},{"id":65397269,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-04-06T20:17:01+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-06 20:17:00","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9266968","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9266968","identity":"rs-9266968","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}