Machine learning for determining accurate outcomes in criminal trials

Abstract Advances in mathematical and computational technologies have brought unique and ground-breaking benefits to diverse fields throughout society (engineering, medicine, economics, etc.). Within legal systems, however, the potential applications of data science and innovative mathematical tools have yet to be embraced with the same ambition. The complex decision-making that is needed for reaching just verdicts is often seen as out of reach for such approaches and, in the case of criminal trials, this inhibits exploration into whether machine learning could have a positive impact. Here, through assigning numerical scores to prosecution and defence evidence, and employing an approach based on dimensionality reduction, we showed that evidence strands presented at historical murder trials could be used to train effective machine-learning algorithms (or models). We tested the evidence quantification approach with the trained model and showed that, through machine learning, criminal cases could be clearly classified (probability >99.9%) as belonging to either a guilty or a not-guilty category. The classification was found to be as expected for all test cases. All guilty test cases that were not wrongful convictions were correctly assigned to the guilty category by our model and, crucially, test cases that were wrongful convictions were correctly assigned to the not-guilty category. This work demonstrated the potential for machine learning to benefit criminal trial decision-making, and should motivate further testing and development of the model and datasets for assisting the judicial process.