Diversity and dermatology through the lens of generative artificial intelligence portrayals: a demographic analysis of race and gender representation

Le résumé fourni par la source

Dermatology remains one of the least diverse medical specialties.1 Recent controversies surrounding diversity, equity, and inclusion initiatives highlight the need to address persistent inequalities within the field.2 In this context, artificial intelligence (AI) presents a unique lens to explore perceptions of diversity. Specifically, text-to-image generative AI is trained on existing image databases to generate outputs3 and may thus reflect existing implicit biases informing visual depictions of major demographic parameters. Consequently, we evaluated gender and race in a sample of AI-generated images of dermatologists to better understand representation in dermatology through the lens of this technology. Five prominent text-to-image generators (DALL-E2, Runway, Midjourney, ImagineAI, and JasperArt) were queried to generate images for the search terms: “dermatologist's face,” “portrait of a dermatologist,” and “dermatologist.” After training on the Chicago Face Database, a previously validated standardized visual database to reduce bias in classification,4 two independent reviewers categorized race and gender for each image. A third reviewer resolved discrepancies. Multiple reviewers categorized each image to reduce inherent biases in visual classification. Images with incomplete or multiple faces were excluded. Race and gender distributions of the AI-generated images were compared to the 2022 Association of American Medical Colleges Physician Specialty Data Report (“AAMC report”) of active dermatologists.5 All statistical analyses were performed in SPSS. In total, 1200 images were generated (240 per AI generator, Figure 1). Of those images, race distribution was significantly different between AI-generated images and the AAMC report (P = 0.006, Table 1). More White (77% vs. 73%) and Asian (19% vs. 15%) dermatologists and fewer Black dermatologists (3.6% vs. 4%) were depicted by AI generators. When AI generators were examined individually, 60% depicted a greater proportion of White dermatologists than the AAMC report, while one (ImagineAI) only depicted White dermatologists. Similarly, 80% of AI generators depicted Black dermatologists at a lower proportion than reported. A significantly higher proportion of women dermatologists was noted among AI-generated images as compared to the AAMC report (61% vs. 52% women, P < 0.001). When AI generators were analyzed individually, this finding held true for DALL-E2 (60% women, P < 0.001) and Midjourney (80% women, P < 0.001). Our analysis highlights an imbalance between dermatologist demographics as represented by generative AI and existing census data. A disproportionate number of dermatologists in AI-generated images were White or Asian and women, with one platform alarmingly depicting only White dermatologists. Fewer Black dermatologists were depicted compared to the AAMC census. Our study supports the need for caution related to the use of generative AI and necessitates further research better to understand its implications for dermatologists and their digital footprint. However, as a reflection of its underlying source imagery, text-to-image generative AI also obviates the biases pervading visual representations of real providers. These biases stem from the training datasets used by AI models, which are often compiled from internet-sourced images and thus can be unrepresentative of all demographic groups and may reinforce existing racial and gender disparities. Consequently, the results of this analysis not only confirm the persistence of inequities in emerging AI representation and support the need for robust systems advocating for diversity. The primary limitation in this analysis is the subjective assessments of race and gender, though we used two raters and a validated image classification system4 to minimize misclassification. Another limitation is that the AAMC report only captures dermatologists in the US, while the AI generators studied were not necessarily engineered to create images based on the US population. However, generative Al has been primarily dominated by US-based companies. Further research should explore generative AI bias in other subspecialties and should explore ways of reducing and mitigating bias in AI datasets.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.