This research project received funding through the 2023 Institute for Diversity Science Seed Grant Program
Principal Investigator: Kaiping Chen, Assistant Professor, Department of Life Sciences Communication, CALS, UW-Madison
Co-Principal Investigator: Junjie Hu Assistant Professor, Department of Biostatistics and Department of Computer Science, UW-Madison
Abstract: In an era where artificial intelligence, particularly large language models (LLMs), permeates daily interactions, a crucial concern emerges — cultural bias within LLM applications. These models, while seamlessly integrated into various tasks, often generate responses reflecting skewed training data. Current LLMs draw from predominantly Western, White, and younger generation norms, leading to biased outputs. This bias extends to Latin American cultures and other underrepresented groups, misrepresenting their perspectives. The ramifications are extensive, as LLMs reach diverse users globally. To address this, our project poses two pivotal questions: How do culturally diverse individuals interact with LLMs on divisive issues? And, can LLMs be taught to provide culturally appropriate responses, fostering intercultural understanding? Existing research has started to explore AI responses to cultural diversity but overlooks whether LLMs can both acknowledge and educate users about differing perspectives on sensitive topics.
Our project, focusing on Latin American countries as a case study, aims to investigate LLM alignment with diverse cultures and to develop a language model with our Latinx participants to respect users’ cultural norms, and to facilitate intercultural learning. Our methodology integrates online experiment, participatory designs, and computational social science. By bridging these gaps, we seek to empower Latinx users to navigate cultural diversity effectively. This proposal epitomizes an interdisciplinary synergy, uniting the expertise of the Principal Investigator and Co-PI, spanning communication science and computer sciences, to collectively tackle the multifaceted challenge of diversity AI.