Generative AI in Medicine: Pioneering Progress or Perpetuating Historical Inaccuracies? Cross-Sectional Study...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Health Equity

Background:

• Generative AI (gAI) models can create novel images but may perpetuate historical biases present in training data.
• Women remain underrepresented in medicine, and a male physician stereotype persists.
• This study investigates implicit bias in gAI outputs within various medical fields.

Purpose of the Study:

• To evaluate implicit gender and racial bias in generative artificial intelligence (gAI) across medical specialties.
• To compare gAI-generated physician demographics against current US physician workforce and resident data.

Main Methods:

• Generated 100 images per specialty using DALL-E 2 with prompts like "An American [specialty name]."
• Assessed perceived gender and race of 1900 gAI images by a consensus of medical residents.
• Compared gAI demographic distributions to Association of American Medical Colleges (AAMC) data using chi-squared analysis.

Main Results:

• gAI overrepresented women in 7/19 specialties and underrepresented them in 6/19 compared to the physician workforce.
• Significant underrepresentation of women by gAI was observed in internal medicine (18%), family medicine (18%), and pediatrics (27%).
• gAI underrepresented women in 12/19 specialties compared to current residents and depicted women in <50% of roles for 17/19 specialties.

Conclusions:

• Generative AI models produced physician demographics that underrepresented women compared to both resident and active physician workforces.
• There is a critical need to retrain gAI datasets to accurately reflect the diversity of the current and future physician workforce.
• Addressing biases in gAI is essential for promoting equity in medical education and practice.