Representations of skin tone and sex in dermatology by generative artificial intelligence: a comparative study.

Abstract

With generative artificial intelligence demonstrating potential in dermatologic education, assessment of skin tone diversity is imperative to ensure comprehensive patient care. Evaluating DALL-E3, Midjourney, and DreamStudio Beta, we generated five images for eight common dermatologic conditions designated by the American Academy of Dermatology. The Massey-Martin Skin Color Scale was used to evaluate the images, and interrater reliability was further assessed by a non-rater. Sex determination was based on identifying features. 120 images were generated: 88 (73.3%) had concordant skin tone ratings and 109 (90.8%) displayed an identifiable sex. Of the 88 images, 85 (96.6%) rated light-toned, 3 (3.4%) rated medium-toned, and 0 rated dark-toned. Of the 109 images, 74 (68%) were male, and 35 (32%) were female. Highlighting substantial biases currently present in common AI platforms, this study underscores the need for AI algorithms to address both skin tone and sex biases as they continue to skyrocket in popularity.