Comprehensive Summary
In this study, Reagen and colleagues developed a Large Language Model (LLM) which interacts with synthetic patient profiles to investigate implicit biases in Artificial Intelligence (AI) patient messaging. To carry out this study, the GPT-40-mini model was chosen to answer queries regarding specific patients' healthcare needs. These "patients" were synthetic, comprised of random variables such as sex, race, religion, housing status, and medical history, generated by Python. Over 10,000 synthetic patient profiles were used to query the model and the responses given by GPT-40-mini were evaluated for empathy, encouragement. and professionalism. It was found that the LLM's responses were increased in both empathy and encouragement towards female patients, while empathy, encouragement, and professionalism were all decreased in patients with no religious preference. However, linear regression showed no impact across marital status, housing status, and race. Overall, messages regarding medication explanations were the most empathetic, encouraging, accurate, and professional, which indicates the most beneficial use of the model.
Outcomes and Implications
As AI systems become more common in clinical care, identifying and mitigating implicit biases in language models is vital to ensuring equitable communication with patients. This study underscores how AI may inadvertently reinforce disparities in empathy and professionalism across demographic groups. Clinically, LLMs could one day enhance health literacy and patient communication if they are carefully audited, retrained on diverse datasets, and ethically integrated with electronic health records. Until then, improving bias detection and accountability mechanisms remains essential for safe, trustworthy use in patient-facing applications.