The thinness of GenAI: body size in relation to the construction of the normate through GenAI image models

Abstract

While generative AI (genAI) image models are increasingly popular, they are not without critique for their biased outputs. Building on assessments of Dall-E’s prejudiced and homogenising production of race, this paper seeks to understand how fat bodies are presented compared to straight-size bodies in 649 images created by nine different, free-to-use genAI image models. The images are examined through critical visual analysis and reflexive thematic analysis. In the first instance, auditing highlights that, if not explicitly prompted to show larger bodies, none of the models create fatness or disability. Secondly, in the outputs with a larger body-size prompt, the models produced images which contravened their own content guidelines, showed fewer positive facial expressions, and had higher rates of mistakes and anomalies compared to images without a body-size prompt. This paper argues that the social imaginaries created through genAI images are foreclosing on difference and forming new normate standards of personhood, which explicitly exclude people who exist in socially deviant bodies.