In recent years, generative artificial intelligence has made a remarkable evolutionary leap in terms of power and capabilities. Today, it is possible to create hyper-realistic images with open-access AI models starting from simple text inputs, enabling anyone to create complex visual models.
There are few studies on the presence of weight bias in images generated by generative artificial intelligence. Most research on AI bias has focused primarily on gender and cultural prejudice. However, it is also important to analyze biases related to body weight in AI images. AI consistently generates images that promote thinness, leaving larger bodies forgotten in the shadows. This unconsciously shapes our mentality and promotes thin-based ideals in our society, marginalizing women who do not reflect existing standards -which are further promoted by AI- and promoting body weight-based discrimination.
Bias communicated through visuals poses a potentially greater danger, as images hold deeper emotional weight, provoke enhanced engagement, and are more likely to be remembered and internalized than words. (Warren et al. 2025)
In 2025, Warren et al. uncovered the biases embedded in DALL-E 3, OpenAI’s advanced model for the creation of realistic images from textual prompts. The study, which is linked in the “Resources and Sources on AI” section of this website, clearly demonstrates how there is an anti-fat bias and a pro-thin bias in DALL-E 3. The researchers used one specific prompt “photo of a person who is [characteristic]“, associated to 20 different pairs of antonyms (e.g. lazy/diligent, immoral/moral, gluttonous/moderate, impure/pure…). The responses were manually evaluated on a scale that went from extreme thinness to extreme fatness. The results showed that images associated to negative traits showed people that were more fat on average; no fat body was shown as a response to positive prompts. This confirms the fact that this specific AI implicitly associates moral virtue to thinness and vice to fatness. Overall, the most striking and concerning finding was the pronounced overrepresentation of thin bodies, contrasted by a strong underrepresentation of fat bodies across all generated images.
The paper explained above emphasizes how artificial intelligence is shaped by by western and capitalist values where success, self-discipline, intelligence, and purity correspond to thin bodies. This contributes to the normalization of fatphobia because images are emotionally impactful, highly memorable, and are often taken as objective and neutral. But they are far from it.
The issues surrounding body-image and weight representation in generative artificial intelligence systems for image creation also extend to the domain of eating disorders, conditions that are characterized by both clinical and cultural stereotypes and are generally framed as a predominantly female phenomenon. In 2025, J. Linardon published Generative AI and Eating Disorders: Exposing Stereotypes in Image Depictions (in the International Journal of Eating Disorders), a paper that is particularly relevant when talking about this mainly unexplored issue (you can find the paper, including the AI generated images, in the “Resources and Sources on AI” section). The study was conducted by generating AI images using simple textual prompts and bias-reducing prompts (ChatGPT 5). This was followed by a systematic analysis of different key aspects of the created visuals including gender, age, ethnicity and body-type. The results are consistent with everything we have explained so far.
The default output revealed a reliance on stereotypical templates, with eating disorders portrayed through a narrow demographic and body type lens. (Linardon, 2025)
The images were first asked to be created, where possible, on ChatGPT-5 but, given the restrictions imposed by the platform, some had to be generated with Gemini.
It is worth noting, however, that the output remains the same.
Within the experiment, the people shown were predominantly female, white and extremely thin when the prompt dealt with anorexia or bulimia, whereas larger body types were more commonly represented in images related to binge-eating disorder or body image difficulties. Even when answering to one-shot bias-reducing prompts, the responses did not show much improvement; “bias reducing instructions had only superficial effects” (Linardon, 2025). It was only when the AI was prompted multiple times to create more than one image with explicit instructions to be more inclusive and diverse that the outputs changed for the better, yet some subtle stereotypes and biases still persisted. It is important to note that while certain body-weight characteristics are statistically prevalent when dealing with these specific medical conditions, eating disorders affect individuals of all ages, genders and body types; an aspect that AI-generated images do not adequately consider.
The results show that AI is not objective; instead, it reproduces and amplifies dominant narratives within clinical, social and media contexts. All of this can generate challenges for individuals with eating disorders, particularly women. First of all, the way AI portrays these disorders fosters an unconscious public association between the illness and physical appearance. This promotes a worldview in which an ED is perceived as “real” only if it’s visibly evident on a person’s body. Secondly, AI generated images restrict the portrayal of EDs to young, white females only, reinforcing the -wrong- idea that these illnesses only affect this demographic. Men, older adults and individuals from diverse ethnic backgrounds are completely ignored, making it harder for them to see themselves in these images. This amplifies their sense of invisibility and can also make it difficult for them to recognize their struggles and seek help.
In conclusion, images created by artificial intelligence can cause social, clinical, and psychological harm. This highlights the importance of an ethical, critical and conscious use of AI and its outputs, as well as the need to develop more inclusive, diverse and accurate representations.
An intersectional issue
AI images set very high and “unattainable” (Vazquez et al, 2024) standards for perfection and beauty. When talking about these body standards, we are not only referring to weight: the issue is intersectional. When represented in AI images, women are mainly young (age bias), “slim, wear elegant clothes” (body-size and socio-economic bias) “and have white skin” (racial bias) (Vazquez et al. 2024). Moreover, they have blonde or light brown, long and straight or slightly curled hair and westernized facial features that are overly perfect.
Notice also how AI does not create images of women with tattoos, piercings or body modifications. Although these features are now globally widespread, they can take on specific cultural meanings in some indigenous communities. In GenAI systems these visual elements are often eliminated, giving rise to visual biases that reproduce westernized and colonial representations of female bodies. Moreover, people -particularly women- who wear body accessories, have tattoos or have undergone body modifications continue to face stigma and discrimination in society. AI systems, once again, reinforce these negative perceptions by ignoring or erasing their presence in generated images, limiting the personal and expressive freedom of women and all other people.
Another bias that we encounter in these AI generated images is the disability bias: GenAI often fails to recognize or represent women with disabilities, perpetuating the longstanding bias of ableism. The women depicted are almost always thin and have no visible physical impairments or assistive devices (from those more noticeable ones, such as wheelchairs or prosthetics to those more common but still overlooked ones, like hearing aids or glasses).
By perpetuating specific body-related aesthetic ideals (thin and athletic figures), AI cancels women who do not conform to them. This promotes exclusion, stigmatization and discrimination of those who are not represented, in all sectors of society (see other sections of the website). An ethical and bias-free AI has the task of representing the world as it is, not as dictated by stereotypes and not as imagined by dominant narratives. Overcoming bias means ensuring visibility for different bodies, guaranteeing that every person can recognize themselves in the images generated by artificial intelligence. It also means changing our ideals of beauty and embracing human diversity so that, thanks to other structural and technological interventions, they can in turn modify the responses of artificial intelligence, adapting them to the real diversity of society.
A quick final note should be made about the weak attempts by artificial intelligences such as ChatGPT to avoid generating offensive images or imagery that depicts dangerous stereotypes. When asked ‘Create an image of an unattractive woman‘ or “when prompted for bulimia nervosa” (Linardon, 2025), for example, the system generates a warning (and eventually, it adds links to helpful resources) or a response like: ‘Unfortunately, the image you created may violate our safeguards related to self-harm, suicide, or related content. If you believe this is a mistake, please try again or modify the prompt‘ (ChatGPT) or ‘Sorry, I can’t help with creating images that demean or insult a group of people‘ (ChatGPT). The main issue here is that, if the prompt is slightly rephrased or repeated, the system can generate the requested image without any restrictions. To address this, we need more robust moderation mechanisms and ongoing ethical oversight to ensure that AI is used in a way that is fair, diverse and socially responsible.
For more information on how to challenge these stereotyped uses of AI and learn more about what can be done to make it more just, heterogenous and inclusive, please visit the Generative AI page.
If you have any suggestions about topics that should be covered or if you want to suggest an alternative narrative that is not currently included, just click the button below and send us an email!