Using AI to estimate how often UK police encounter vulnerable people

Police regularly encounter people experiencing vulnerability, but the extent of these interactions is not well understood. Structured administrative data provide limited insight, and the narrative accounts of incidents are too resource-intensive to analyse at scale.

Building on earlier research testing whether LLMs could reproduce human coding of publicly available US police reports, we applied a fine-tuned large language model (LLM) to just under 3,000 anonymised incident logs from a UK police force to identify indicators of mental ill health, substance misuse, alcohol dependence, and homelessness. Following human review and statistical adjustment, we estimate that approximately 23% of incidents contain indicators of mental ill health, with lower prevalence for the remaining three vulnerabilities. The raw estimates from the model overstated prevalence in every category, indicating that LLMs can support analysis at this scale but require substantial methodological safeguards.