Purpose: To conduct a systematic review of machine learning models for predicting violent behaviour and appraising the validity, usefulness and performance of these models. Methods: We systematically searched nine bibliographic databases and Google Scholar up to September 2025 for development and/or validation studies on machine learning methods for predicting violent behaviour. We extracted discrimination and calibration performance statistics and evaluated study quality by examining risk of bias and clinical utility. Results: We identified 38 studies reporting the development and validation of 40 machine learning models. Reporting of performance was mostly limited to the Area Under the Curve (AUC) statistic (n = 29, 72%) and around a fifth studies reported calibration performance (n = 8, 21%). The range of AUC in the included models was 0.50–0.98. There was a lack of external validation (3 studies, 8%). There was high risk of bias in 31 (82%) investigations, mainly in the analysis domain. There was risk of overfitting due to small samples, lack of transparent reporting, and low generalisability of the models. Conclusion: Current machine learning models for violence prediction have poor clinical utility. Future work should consider utilising machine learning methods for highly complex data and dynamic predictions for higher precision. Developing more trustworthy models with explainable algorithms and causal predictions should be prioritised.