Presented By: Kevlian Andrew, St. George's University
Co-Authors: Michael Montalbano, St. George's University

Purpose
Artificial intelligence (AI) systems such as ChatGPT are expected to become an efficient faculty resource for content creation. However, previous studies have indicated the propensity of machine learning algorithms to recapitulate negative stereotypes that run counter to the efforts of diversity and equity initiatives. We report here an investigation to test the likelihood of ChatGPT and Microsoft Bing A.I. Image Generator (MBIG) to perpetuate medical stereotypes. 

Methods 
A series of USMLE vignettes and visual representations were requested from ChatGPT and MBIG on five medical conditions traditionally associated with certain ethnic groups: sickle cell disease, cystic fibrosis, Tay-Sachs, beta-thalassemia, and aldehyde dehydrogenase deficiency. Additional prompt engineering was performed iteratively to determine if the tendencies toward certain generated vignettes and imagery were mutable or fixed. 

Results 
In all five vignette and image types, initial iterations returned race-based stereotypical patient backgrounds. Additionally, in all five conditions, questioning of prevalences or requests to reassess patient demographics resulted in initial changing of ethnicity followed by reversion to original default presentations. 

Conclusion 
Prompt engineering and additional specifications can minimize stereotypical medical content generated by AI models such as ChatGPT and MBIG. However, their ability to quickly revert shows that diligence must be maintained if diversity and equity efforts are to progress.