Large Language Models in Civic Education on the Supervision and Risk Assessment of Public Works

Abstract

The Public Administration spends an estimated 13 trillion USD annually worldwide, of which approximately 20% is allocated to public works. Despite strict rules, unfinished works for legal reasons, including corruption, are not atypical, negatively impacting the region’s economy, culture, and society. Civic awareness about this problem may help reduce such losses. This study investigates the use of Large Language Models (LLM) and Retrieval-Augmented Generation (RAG) to support civic education on risks in public works. While LLMs interpret and create human language, RAGs combine text production with access to other external data, allowing contextualized responses. Here, we evaluate how these technologies can facilitate the population’s understanding of technical information about public works. To this end, we initially create and evaluate 4 Machine Learning models for risk prediction of public work failure, using data from real public works. We provide a failure estimate for each contracted work based on the most efficient model. These data and others related to government development and risk processes are accessed and presented to the user through a web support system. Tests with 35 participants indicate a significant improvement in citizens ability to understand complex aspects related to risks and contracts of public works.