Current deep learning methods are known to fail unexpectedly and catastrophically for inputs that are unlike the training data. One way of coping with this brittleness is to incorporate a human into the learning loop. One example is the learning to defer framework, where a rejection model decides if an input is passed to the model or deferred to a human (or some other backup system). In this project, we propose to investigate the learning-to-defer framework's ability to handle and cope with distribution shift and out-of-distribution (OOD) detection.