Author: Daha Pavlovic
Supervisor: Julia Neidhardt, Co-Supervisor: Bernhard Krüpl-Sypien
Abstract
A significant use case for large language models (LLMs) is the provision of medical advice, including diagnostic suggestions, treatment plans and healthcare recommendations. As individuals increasingly rely on LLMs for initial medical consultations either before visiting a healthcare professional or, in some cases, instead of seeking professional care, the dependency on these AI systems continues to increase. This raises concerns about the potential biases embedded in model outputs, particularly biases related to race. Bias in medical recommendations can have serious real-world consequences. In particular, unequal treatment recommendations based on a patient’s race may worsen existing healthcare disparities, lead to inappropriate medical advice, and ultimately affect patient outcomes. Therefore, ensuring that LLMs provide unbiased medical guidance is needed for both the safety and trustworthiness of AI in healthcare. This thesis aims to draw attention to the strong need for mechanisms that can detect, measure and mitigate racial bias in the medical applications of LLMs. The goal is to determine how fine-tuning and prompting, as mechanisms for shaping model behaviour, influence these biases, and whether one mechanism is more effective than the other at reducing biases, ultimately contributing to more inclusive healthcare recommendations.