Demographic Representation in 3 Leading Artificial Intelligence Text-to-Image Generators

Area of Science:

• Artificial Intelligence
• Medical Professional Demographics
• Societal Bias in AI

Background:

• Artificial intelligence (AI) text-to-image generators are increasingly sophisticated.
• Concerns exist regarding the perpetuation of societal biases, including profession-based stereotypes, by these AI models.
• Understanding AI's representation of demographic groups in specialized professions is crucial.

Purpose of the Study:

• To evaluate the demographic accuracy of surgeon representation in three leading AI text-to-image models.
• To compare AI-generated images of surgeons against real-world demographic data of attending surgeons and trainees.
• To identify potential biases in AI's depiction of race and gender in surgical specialties.

Main Methods:

• A cross-sectional study analyzing 2400 AI-generated images across 8 surgical specialties from 3 popular text-to-image models.
• Seven independent reviewers categorized AI-produced images based on race (White vs. non-White) and gender (female vs. male).
• Demographic data for comparison were sourced from the Association of American Medical Colleges and the American Medical Association.

Main Results:

• Two AI models generated images with over 98% White and male surgeons, significantly underrepresenting non-White and female individuals compared to real-world data.
• While one model (DALL-E 2) showed demographic parity with attending surgeons, all three models underestimated trainee representation.
• Geographic-based prompts increased non-White surgeon representation in some instances but did not significantly affect female representation.

Conclusions:

• Leading AI text-to-image generators currently amplify societal biases, particularly concerning the underrepresentation of female and non-White surgeons.
• All evaluated models failed to accurately represent the demographics of surgical trainees.
• The findings underscore the urgent need for AI developers to implement guardrails and feedback mechanisms to mitigate the amplification of professional stereotypes.