LLMs Exhibit Place-Based Stigma in Urban Safety Judgments

This research investigates whether large language models (LLMs) used for urban safety assessments rely on actual risk factors or on place-based stigma associated with neighborhood names. The study probed seven instruct-tuned LLMs across 186 neighborhoods in Los Angeles and Chicago, using violent crime and demographic data. It tested three conditions: coordinates-only, name-only, and name-plus-coordinates. The findings reveal that LLM safety ratings are largely flat when only coordinates are provided, indicating that neighborhood names are the primary driver of variation in judgments. These names, while moderately correlated with violent crime, also significantly lower safety ratings for neighborhoods with higher proportions of locally dominant marginalized groups (e.g., Black residents in Chicago, Hispanic residents in Los Angeles). This demographic bias persists even after controlling for crime and income, particularly in Los Angeles where demographics and crime are more separable. Furthermore, the study observed that models with better geographic knowledge tend to apply more demographic stereotypes. While removing neighborhood names reduces bias, it also diminishes accuracy, as names carry both genuine crime signals and stereotypes. This highlights a critical challenge for deploying LLMs in decision-support roles where fairness and accuracy are paramount.