Representation of Demographics in Otolaryngology by Artificial Intelligence Text-to-Image Platforms

Abstract

Objective

Artificial intelligence (AI) text-to-image generators have a propensity to reflect stereotypes. This study investigates the perception of race and gender of AI-generated portraits of otolaryngologists, evaluating their accuracy against workforce demographics and whether they amplify existing social biases.

Methods

Three text-to-image platforms (DALL-E3, Runway, Midjourney) were prompted to generate portrait photos of otolaryngologists based on 29 categories, including personality traits, fellowship, and academic rank. 580 portrait photos were made per platform. Two reviewers characterized the gender and race of the 1740 portraits. Statistical analysis compared the demographics of AI outputs to existing demographic information.

Results

Of the 1740 AI-generated otolaryngologists generated, 88% of images were labeled as White, 4% Black, 6% Asian, 2% Indeterminate/Other race, 88% male, and 12% female. Across academic rank, the representation of White individuals was 97% (department chairs), 90% (program directors), 93% (professors), and 78% (residents). Male representation ranged from 90% (department chairs), 75% (program directors), 100% (professors), and 87% (residents). Runway produced more images of male (89% vs. 88% vs. 85%, p = 0.043) and White (92% vs. 88% vs. 80%, p < 0.001) otolaryngologists than DALL-E3 and Midjourney, respectively.

Conclusion

Text-to-image platforms demonstrated racial and gender biases, with notable differences compared to actual demographics. These platforms often underrepresented females and racial minority groups and overrepresented White males. These disparities underscore the need for the awareness of biases in AI, especially as these tools become more integrated into patient-facing platforms. Left unchecked, these biases risk marginalizing minority populations and reinforcing societal stereotypes.

Level of Evidence

4.