Study Finds Place-Based Stigma in LLM Urban Safety Judgments

Summary

A new arXiv study finds that neighborhood names can strongly shape how large language models assess urban safety. Testing seven instruction-tuned models across 186 Los Angeles and Chicago neighborhoods, researchers found that names explained most rating variation, while coordinates produced nearly flat ratings for six models. Safety scores fell more for neighborhoods with higher shares of locally marginalized groups, and the effect persisted after controlling for crime and income in Los Angeles. Removing names may reduce bias, but it can also remove useful information about actual crime risk.