This paper addresses the challenges of detecting misclassified and out-of-distribution examples in neural networks. It introduces a simple baseline that leverages softmax probabilities to distinguish between correctly and incorrectly classified examples, as well as between in-distribution and out-of-distribution data. The baseline is evaluated across various tasks in computer vision, natural language processing, and automatic speech recognition, demonstrating its effectiveness. However, the paper also highlights that the softmax prediction probabilities can be misleading as confidence estimates, and proposes an auxiliary decoder-based method to improve detection performance. This method reconstructs inputs and uses the reconstruction quality to determine if an example is abnormal. The paper concludes by discussing future research directions, emphasizing the importance of developing more reliable methods for error and out-of-distribution detection.This paper addresses the challenges of detecting misclassified and out-of-distribution examples in neural networks. It introduces a simple baseline that leverages softmax probabilities to distinguish between correctly and incorrectly classified examples, as well as between in-distribution and out-of-distribution data. The baseline is evaluated across various tasks in computer vision, natural language processing, and automatic speech recognition, demonstrating its effectiveness. However, the paper also highlights that the softmax prediction probabilities can be misleading as confidence estimates, and proposes an auxiliary decoder-based method to improve detection performance. This method reconstructs inputs and uses the reconstruction quality to determine if an example is abnormal. The paper concludes by discussing future research directions, emphasizing the importance of developing more reliable methods for error and out-of-distribution detection.