AI Based Smart Policing System Using Large Language Models

Authors: Bhagyashri R. Kasar, Prashant M. Yawalkar, R. P. Dahake

Abstract: The need for more sophisticated policing methods is growing as crime rates rise and crime data processing becomes increasingly difficult. In order to increase law enforcement effectiveness, this has prompted the adoption of smart policing systems and predictive policing techniques that make use of artificial intelligence (AI). Among the many AI technologies used, machine learning (ML) is regarded as a powerful instrument for identifying patterns, analyzing crime trends, and predicting criminal activity. LLMs generative AI has recently gained widespread recognition in sectors like agriculture, healthcare, law, and finance. However, the LLM's potential in smart policing and crime prediction is still mostly unrealized. By creating a framework that integrates the most advanced LLMs—BART, GPT-3, and GPT-4—this study seeks to bridge this gap and improve analysis and crime prediction. Using actual crime datasets from San Francisco and Los Angeles, the framework's performance is evaluated using zero-shot prompting, few-shot prompting, and fine-tuning techniques. In the majority of the experimental setups, a comparative analysis between the LLM-based methods and the conventional ML models revealed that GPT models performed better in terms of crime classification. The study offers future paths for AI-driven police and demonstrates how LLMs have the ability to transform contemporary crime analysis.

DOI: https://doi.org/10.5281/zenodo.17962534