Hot take: Data augmentation is a double-edged sword. Balancing in-distribution and out-of-distribution techniques is key to building models.