Using large language models (LLMs) to support police problem-solving: Considering community concerns at homeless encampments

Police collect many kinds of data and are expected to analyze it for problem-solving. But there’s insufficient capacity to analyze all data. We show how police can apply large language models (LLMs) to unstructured data for problem-solving. We do this by applying LLMs to unstructured residents’ reports of homeless camping. LLMs are prompted to code resident concerns from 2021-2024 in Boulder, Colorado. A total of 10,286 reports were coded to identify 10 kind of problems. Substantial agreement between LLM coding was found (Cohen’s K = .67). Longer reports more frequently had concerns. One percent of reporters submitted 31% of all reports. The LLMs described camps as having any concern (58%), trash (34%), strange behaviors (10%), stolen bicycles (10%), drugs (8%), stolen property (7%), human feces (7%), shopping carts (5%), fires (4%), propane tanks (3%), and weapons (1%). This study demonstrates that LLMs have the potential to aid police problem solvers by organizing unstructured data that often goes unused in problem-solving.