Disclosed are various embodiments for using contrarian machine learning models to compensate for selection bias. Both a primary machine learning model and a contrarian machine learning model may be trained for selecting sets of items based at least in part on the same training data. However, the contrarian machine learning model is specially trained to avoid selecting items that are selected by the primary machine learning model. Items selected by the primary model and items selected by the contrarian model are presented to users as recommendations. Both models are updated based at least in part on user selections of items. Ultimately, the use of the contrarian model avoids causing the primary model to degenerate to picking random items due to reinforcement resulting from a bias in favor of selecting items that have been recommended.