Machine learning in legal decision-making: analysis of judicial and algorithmic reasoning in road homicide cases

Abstract Crime scene analysis requires the evaluation of multiple factors to determine a suspect’s guilt, a process that can be lengthy and costly. The integration of Artificial Intelligence (AI) into the judicial system is emerging as an opportunity to improve the efficiency of investigations and legal decision-making. In this study, we propose a Machine Learning (ML)-based methodology to support the assessment of road homicide cases under Italian law. Our approach employs a Large Language Model (LLM) to extract 51 features from crime scene descriptions automatically. Four ML models are then analyzed: (RF), (GBM), (DT), and (LR). We evaluated the performance of these models on a dataset of 100 road homicide court rulings in Italy, achieving 95% accuracy in crime classification. The validation was conducted by comparing the model outputs with a legal ranking established by four legal experts, allowing us to verify the consistency of the algorithmic predictions with human legal reasoning. The results indicate that the shows the highest correlation with legal evaluations ( $$\rho = 0.857$$ ρ = 0.857 , $$\tau = 0.714$$ τ = 0.714 , p values $$< 0.05$$ < 0.05 ), while performed the worst. This study highlights the potential of AI in legal decision support, emphasizing the need to ensure transparency and bias mitigation to comply with European Union Regulations while maintaining human judgment as the central authority in legal proceedings.