In the domain of Visual Question Answering (VQA), studies have shown\nimprovement in users' mental model of the VQA system when they are exposed to\nexamples of how these systems answer certain Image-Question (IQ) pairs. In this\nwork, we show that showing controlled counterfactual image-question examples\nare more effective at improving the mental model of users as compared to simply\nshowing random examples. We compare a generative approach and a retrieval-based\napproach to show counterfactual examples. We use recent advances in generative\nadversarial networks (GANs) to generate counterfactual images by deleting and\ninpainting certain regions of interest in the image. We then expose users to\nchanges in the VQA system's answer on those altered images. To select the\nregion of interest for inpainting, we experiment with using both\nhuman-annotated attention maps and a fully automatic method that uses the VQA\nsystem's attention values. Finally, we test the user's mental model by asking\nthem to predict the model's performance on a test counterfactual image. We note\nan overall improvement in users' accuracy to predict answer change when shown\ncounterfactual explanations. While realistic retrieved counterfactuals\nobviously are the most effective at improving the mental model, we show that a\ngenerative approach can also be equally effective.\n