AI12 Bad skin, bad guy? Dermatological tropes in artificial intelligence-generated characters

Abstract

Negative dermatological features such as scarring, alopecia and pigmentary changes have traditionally been used in media to signify villainous characters, contributing to stigma around visible skin differences. As artificial intelligence (AI) is increasingly used to generate fictional characters, there is concern that these systems may reproduce existing societal biases. The aim of this study was to explore whether AI-generated characters reinforce negative dermatological stereotypes by disproportionately assigning visible skin features to villains compared with heroes. We conducted an exploratory qualitative content analysis using three widely used AI platforms. We used standardized prompts for these AI platforms to generate images of 20 heroes and 20 villains (total n = 120). Characters were assessed for predefined dermatological features, including alopecia, wrinkles, scars, hyperpigmentation, vitiligo and acne or blemishes, based on prior literature. Features were coded as present or absent and compared between heroes and villains using descriptive analysis to identify patterns of stereo­typing. Villains in AI-generated images showed significantly higher rates of four facial attributes than heroes: wrinkles (60% vs. 27%), hyperpigmentation (48% vs. 5%), alopecia (25% vs. 10%) and scars (23% vs. 3%) (all P < 0.05). These differences are consistent across ChatGPT, Gemini and Canva; Canva yields the highest absolute levels, Gemini intermediate and ChatGPT the lowest. Vitiligo and acne/blemishes are rare and not differentiating. Male characters exhibit significantly more wrinkles, hyperpigmentation and alopecia than female characters, with the gender gap largest among villains. Facial-impairment attributes increased with age. AI-generated imagery encodes clear visual biases that reinforce negative dermatological tropes: wrinkles, hyperpigmentation, alopecia and scars are used as shorthand for villainy, perpetuating associations of visible skin differences with moral failing, otherness and ageing. These attributes are disproportionately applied to male and older characters and persist across platforms, indicating systemic model and dataset bias.