Mitigating Bias in Large Language Models Through Culturally-Relevant LLMs
Malur Narayan, John Pasmore, Elton Sampaio, Raghavan Vijay, Sayan Maity, Gabriella Waters, Ayanna Howard · 2025
With the proliferation of Large Language Models (LLMs) and its use across sectors that impact our civil liberties, the biases present in these language-based applications can propagate and amplify social inequalities and harmful stereotypes. In particular, racial and cultural biases in LLMs not only undermine the fairness and reliability of these systems, but also pose significant ethical concerns, further affecting marginalized communities. To address these concerns, this paper discusses the development of a LLM for mitigating bias through the integration of diverse data sources and culturally relevant material that has often been overlooked in the construction of mainstream AI models. We compare the performance of ChatGPT 3.5 and the LLM, which was incrementally trained on Black history and culture, when tasked with answering a set of prompts that cover a range of news topics. Using a bias assessment score, we assessed each model's ability to handle the task without introducing racial or cultural biases into the news topic summaries. Through empirical studies, our approach not only demonstrates the feasibility of detecting and measuring racial and cultural biases but also offers a scalable solution for creating more equitable and reliable AI technologies.