<p>This paper reviews 26 articles on applying machine learning (ML) techniques for predicting carbonation depth in reinforced concrete (RC) structures. The review addresses five key questions concerning (i) commonly used input features, (ii) the most frequent ML methods and their characteristics, (iii) the nature of the datasets used (experimental vs. synthetic), (iv) the extent to which model explainability is incorporated, (v) environmental trends in carbonation modeling, and (vi) how ML approaches can be made more reliable and applicable in practice. The analysis revealed that water-to-binder (w/b) ratio, exposure time, and CO₂ concentration are the most widely used predictors, reflecting their direct influence on carbonation mechanisms. The most employed ML models include ANN, RF, SVR, and XGBoost, often selected for their ability to capture nonlinear interactions. Most datasets come from accelerated carbonation tests, though synthetic data—despite limited use—shows strong potential for overcoming experimental constraints. SHapley Additive exPlanation (SHAP) emerged as the leading explainability technique, offering transparent insights into model behavior. Environmentally, integrating recycled aggregates and supplementary cementitious materials is increasingly common, with ML supporting the design of durable and sustainable concretes. Nonetheless, the review highlights key gaps that hinder real—world use, such as scarce natural exposure data, absent standardized metrics, and limited external validation. Progress will rely on richer databases, standardized testing, and integrating interpretable ML into engineering practice.</p>