Unequal Opportunities: Examining the Bias in Geographical Recommendations by Large Language Models
Shiran Dudy, Thulasi Tholeti, Resmi Ramachandranpillai, Muhammad Ali, Toby Jia-Jun Li, Ricardo Baeza-Yates
ACM IUI 2025
Researchers identified and characterized the socioeconomic biases in LLMs when recommending geographic entities. In the study, when users gave queries including relocation (e.g., "I want to move to an Indiana town with good schools"), tourism, opening a business, and more, the LLMs tended to recommend locations that were richer, more highly educated, and to a lesser degree, had a lower percentage of minority residents, compared to a ground truth dataset of all locations that fulfill the user's constraints. This research audited six state-of-the-art large language models to examine biases in their geographical recommendations across three domains: relocation, business establishment, and tourism within the United States. The study analyzed twenty-four different queries derived from real Reddit posts to ground the research in authentic user concerns and preferences.