Large language models (LLMs) have exploded in usage, but have significant problems with social biases, misinformation, security, and more. This project examines dialect discrimination, specifically in how LLMs codify stereotypes and discriminatory decision-making against users of African American English. We propose a new method that uses biased character trait associations in large language models to decrease dialect bias in downstream use cases, including in decision-making scenarios.

I am excited to continue my forays into machine learning, especially in safety and fairness, through the SuperUROP program. I feel best prepared for this project by the class 6.S977, a prior UROP with my supervisor, and two years in the SERC Scholars program. I am grateful for this opportunity, and am eager to use this experience to continue working in the ML space!