Construction sites remain high-risk environments, with thousands of accidents occurring each year. Although large amounts of accident records exist globally, much of this information is stored as unstructured free-text comments, limiting its usefulness for systematic analysis. This study addresses this gap by developing an artificial intelligence (AI)–based methodol- ogy to transform unstructured accident descriptions into structured insights. Approximately 12,000 construction site accidents from multiple countries were analyzed, initially written in nearly 20 languages and containing minimal structured data. The aim of this research is to automatically classify accidents into meaningful root causes using large language models (LLMs) and to leverage these classifications to improve understanding of accident mech- anisms and guide safety interventions. LLM-enhanced processing enabled the extraction of contextual variables and the identification of prioritized root causes for each case. The analysis was conducted across several levels, from descriptive comparisons (e.g., severity by country or season) to more advanced network analysis of chained root causes. A focused case study of a single site with more than 1200 accidents provided deeper insights into re- curring failures and the prevalence of worker-attributed causes. Validation through manual classification of 150 accidents showed a 100% match with the LLM-generated categories, demonstrating the reliability of the proposed approach. Findings reveal variations in acci- dent patterns across economic contexts and raise concerns about the frequent attribution of responsibility to workers. Based on the extracted root causes, tailored policy recommenda- tions are proposed to support more effective safety strategies. This work contributes to a deeper, data-driven understanding of construction site risks and provides a scalable method to uncover systemic causes behind accidents. • Develop an analytics framework that converts narrative accident reports into structured inputs for safety decisions. • Classify large-scale construction accidents into actionable root causes to support data-driven risk management. • Analyze chains of interconnected accident causes to reveal escalation patterns affecting decision priorities. • Integrate accident severity into analytical models to improve the prioritization of safety interventions. • Uncover systematic bias in accident attribution to inform more balanced organizational safety decisions.
Omari et al. (Wed,) studied this question.