Large Language Models in Otolaryngology Residency Admissions: A Random Sampling Analysis

Area of Science:

• Medical Education
• Artificial Intelligence
• Health Equity

Background:

• Residency selection committees (RSCs) in otolaryngology aim for unbiased evaluation.
• Artificial intelligence (AI) is increasingly explored for applications in medical education and selection.

Purpose of the Study:

• To investigate demographic bias in AI-driven simulations of otolaryngology residency selection.
• To assess if AI models like GPT-4 and GPT-4o exhibit bias similar to human reviewers.

Main Methods:

• Simulated RSC members with diverse demographics used an API to interact with GPT-4 and GPT-4o.
• 10 applicants with identical qualifications but varied demographics were evaluated.
• 1000 simulations were run per RSC, with chi-square tests analyzing bias and GPT-4o providing rationales.

Main Results:

• Simulated RSCs demonstrated significant racial, gender, and sexual orientation bias (p < 0.05).
• RSCs showed preference for applicants of shared demographics; Asian male applicants had lowest selection rates.
• GPT-4o favored Black female and LGBTQIA+ applicants, citing inclusivity in over 95% of decisions.

Conclusions:

• Publicly available large language models (LLMs) may introduce substantial bias into otolaryngology residency selection.
• The evolving nature and potential for bias in LLMs require careful consideration and mitigation strategies.
• Minimizing LLM bias is crucial for fostering a diverse and representative otolaryngology workforce.