Skin tone bias in generative artificial intelligence for use in medical education

Background Image generative artificial intelligence could be useful to medical educators, particularly in the disciplines of anatomy and dermatology. Medical textbooks have been noted to contain a paucity of images with subjects of a darker skin tone.1 This study aimed to test if the same lack of diversity is also present in medical images generated by artificial intelligence.

Methods A prompt was given to two Artificial Intelligence image generation models (Dall-E and Midjourney) to generate images (n = 200) of people with psoriasis. Three researchers separately rated each image using the validated Massey-Martin skin tone rating scale.2 The median skin tone rating was taken to represent each image. A goodness-of-fit test (Pearson's Chi-squared) was undertaken to compare the distribution of skin tones in the AI- generated images to an expected distribution of skin tones based on the American National Election Survey Time series 2012 study.3

Results Pearson's Chi-squared goodness-of-fit analysis showed a statistically significant difference existed between AI-generated skin tones and skin tones that might be encountered in society (p < 0.001). Educators who opt to use generative AI should be aware of its significant bias towards lighter toned skin. Further work should examine whether more sophisticated prompts can overcome this bias to create images which reflect the expected distribution of skin tones to be representative of the desired population. Other work should be undertaken to establish whether similar biases exist elsewhere in generative AI.