Seeing Is Believing? Exploring Gender Bias in Artificial Intelligence Imagery of Specialty Doctors.

BACKGROUND: In medicine and medical education, women are disproportionately affected by gender bias. Artificial intelligence (AI) is increasingly being employed in medical education. As gender bias exists within AI, there is a risk of reinforcing gender stereotypes if AI is used to generate images of medical professionals. We examined whether the gender distribution of doctors seen in AI-generated images was representative of UK specialty trainee doctors. METHODS: Free-to-use AI text-to-image generators were used to create 1200 images across 30 specialties. NHS England recruitment data provided figures on gender. Specialties accounting for < 0.25% of overall recruitment were excluded as small numbers precluded meaningful analysis. Each image was independently reviewed by both authors and classified (male/female/not-classifiable). Any disagreement was resolved by discussion. 'Not-classifiable' images were removed from analysis. Gender distribution between the AI images and recruitment data was compared (chi-squared test, significance p < 0.05). FINDINGS: There was a significantly higher proportion of males in the AI-generated images compared to NHS specialty data (82% vs. 47%; p < 0.0001). Notably, both AI tools created no images of female general practitioners, orthopaedic surgeons or urologists. Conversely females were overrepresented as dermatologists, obstetricians and gynaecologists and plastic surgeons. CONCLUSION: The finding of representational and presentational gender bias in AI-generated images of doctors is consequential because 'visual culture' within medical school, and beyond, matters. We contend that healthcare educators ought to employ caution when using AI and consider developing guidance on responsible use of AI imagery; otherwise, they risk perpetuating, rather than challenging, harmful gender stereotypes about medical career pathways.