A processor receives requests for content items and identifies a first subset of machine learning (ML) models that satisfy a reliability criterion and a second subset of ML models that fail to satisfy the reliability criterion, wherein each ML model is associated with a respective content template and is trained to output a probability that a target associated with an input set of characteristics would perform a target action responsive to being presented with a content item generated based on the respective associated content template. For each request in a first group, the processor inputs the respective set of characteristics associated with the request into each ML model of the first subset, selects a content template, and generates a content item based on the selected content template. For each request in the second group, the processor generates a content item based on a content template associated with the second subset.