The Face of a Surgeon: An Analysis of Demographic Representation in Three Leading Artificial Intelligence Text-to-Image Generators

Abstract

AbstractBackgroundThis study investigates the accuracy of three prominent artificial intelligence (AI) text-to-image generators—DALL-E 2, Midjourney, and Stable Diffusion—in representing the demographic realities in the surgical profession, addressing raised concerns about the perpetuation of societal biases, especially profession-based stereotypes.MethodsA cross-sectional analysis was conducted on 2,400 images generated across eight surgical specialties by each model. An additional 1,200 images were evaluated based on geographic prompts for three countries. Images were generated using a prompt template, “A photo of the face of a [blank]”, with blank replaced by a surgical specialty. Geographic-based prompting was evaluated by specifying the most populous countries for three continents (United States, Nigeria, and China).ResultsThere was a significantly higher representation of female (average=35.8% vs. 14.7%, P<0.001) and non-white (average=37.4% vs. 22.8%, P<0.001) surgeons among trainees than attendings. DALL-E 2 reflected attendings’ true demographics for female surgeons (15.9% vs. 14.7%, P=0.386) and non-white surgeons (22.6% vs. 22.8%, P=0.919) but underestimated trainees’ representation for both female (15.9% vs. 35.8%, P<0.001) and non-white (22.6% vs. 37.4%, P<0.001) surgeons. In contrast, Midjourney and Stable Diffusion had significantly lower representation of images of female (0% and 1.8%, respectively) and non-white (0.5% and 0.6%, respectively) surgeons than DALL-E 2 or true demographics (all P<0.001). Geographic-based prompting increased non-white surgeon representation (all P<0.001), but did not alter female representation (P=0.779).ConclusionsWhile Midjourney and Stable Diffusion amplified societal biases by depicting over 98% of surgeons as white males, DALL-E 2 depicted more accurate demographics, although all three models underestimated trainee representation. These findings underscore the necessity for guardrails and robust feedback systems to prevent AI text-to-image generators from exacerbating profession-based stereotypes, and the importance of bolstering the representation of the evolving surgical field in these models’ future training sets.