LLM Study Reveals Social Biases Against Homelessness in Online and Offline Discourse

A recent study published on arXiv (2508.13187) introduces the first multi-domain corpus aimed at identifying social biases towards individuals experiencing homelessness (PEH). Released as a replace-cross announcement, the research tackles the ongoing issue of homelessness, which impacted more than 876,000 people in the U.S. in 2025. The authors emphasize that social bias significantly hinders efforts to combat homelessness, affecting public attitudes and policy decisions. They examined both online media and offline city council discussions to uncover these biases. The corpus features a 16-category multi-label taxonomy, including a gold-standard set of 1,698 items annotated by trained raters and 48,389 texts labeled by GPT-4.1. Data sources spanned Reddit, X (formerly Twitter), news outlets, and city council transcripts from ten U.S. cities between 2015 and 2025. The study evaluated six prompted LLMs against the gold-standard set, providing F1 scores alongside prevalence-gap audits. Results show that while LLMs can detect biases, their calibration needs enhancement, offering crucial insights for addressing social biases against PEH and influencing public policy and media portrayal.