Large language model‑based analysis and synthesis of multidimensional safety data in US cities

• In Event: Activist Scholarship

Abstract

The public debate on crime is, at its core, a dialogue on elevating community safety. By focusing on crime, policy solutions often skew towards punitive and deterrent measures. Rather than relying solely on crime statistics—data marred by historical inequities and biases—we embrace a broader conception of safety. Our approach redefines “safety data” to encompass not only crime records but also the social determinants of health.
High‑dimensional safety data in US cities come from myriad sources, including the Census, government agencies, NGOs, and research institutions. Yet, synthesizing these disparate streams into actionable insights poses unique challenges. By consolidating crime and social determinants data from an initial selection of eleven US cities into a unified platform, we have begun to address these challenges. Here we explore a theoretical and modeling approach to multidimensional safety that leverages large language models (LLMs) for the qualitative coding of policy texts and news media reports, feeding into a quantitative model that integrates structured data. The output—a composite safety score and scorecard for each city—serves to highlight how best to allocate resources to improve safety. With this approach, disparate data can be synthesized into actionable policy change, digestible by policymakers, journalists, and the public.

Authors

• Andrew David Zaharia, Campaign Zero

• Mikelina Belaineh, Campaign Zero