Identification of individuals at risk of suicide-related outcomes (ideation, attempt, and death) is critical to the prevention of suicidal behaviors. Machine learning (ML) has generated considerable excitement for its potential to improve suicide risk prediction. ML algorithms can adapt flexibly to large, high-dimensional, and noisy data. Compared to classical statistical methods, ML is more effective at handling non-linearities and interactions among many variables. Numerous proof-of-concept studies attest to ML’s theoretical promise in suicidology, but methodological flaws are pervasive. Validation of prediction models in independent data—a critical next step toward clinical implementation—rarely occurs. Systematic reviews of ML-based prediction models in psychiatry, inclusive of suicide prediction, have found that most studies suffer high risk of bias. Reasons include poor quality and biased data, failure to control model complexity, data leakage, improper (or lack of) cross-validation, and use of inappropriate evaluation metrics, among others. The ML model development process is vulnerable to human inductive biases, which can inflate model performance estimates due to unintentional errors or deliberate “gaming” for publication. Published ML models are rife with overfitting and overoptimism, largely unchecked due to insufficient peer review. The field can no longer ignore methodological issues for the sake of novelty and expect safe, ethnical, or meaningful progress toward clinical implementation. This presentation will provide attendees with an overview of critical issues in the application of ML to suicide prediction.