This paper investigates biases in generated image detection datasets, particularly focusing on JPEG compression and image size. The authors demonstrate that many datasets for AI-generated image detection contain biases related to JPEG compression and image size, which are subsequently used by detectors during inference. Using the GenImage dataset, they show that detectors indeed learn from these undesired factors. Removing these biases significantly enhances cross-generator performance, achieving state-of-the-art results on GenImage and increasing the average accuracy by more than 11 percentage points for ResNet50 and Swin-T detectors. Detectors become more robust against distortions due to now learning the actual task of detecting generation-specific artifacts. The GenImage dataset is one of the largest and most diverse datasets for generated image detection. It includes natural images from the ImageNet1k dataset and approximately an equal number of generated images from various generators. However, the evaluation of a detector based on the raw GenImage Benchmark is not reliable since JPEG compression and image size biases are used by detectors during training. The paper highlights that the use of JPEG compression and image size biases in training data can lead to detectors learning to distinguish between natural and generated images based on these factors, rather than the generation-specific artifacts. This can result in overestimation of detection performance and hinder the adaptability of detectors to real-world scenarios where such biases are absent. The authors conducted experiments to assess whether detectors trained on datasets containing such biases inadvertently acquire information from these undesirable variables. They found that removing these biases significantly improves the robustness and cross-generator performance of detectors. Specifically, they found that detectors trained on a constrained dataset with JPEG compression and size biases removed showed significantly improved performance compared to those trained on the raw GenImage dataset. The paper concludes that biases in generated image detection datasets can significantly affect the performance of detectors, leading to misjudgments in real-world scenarios. By imposing constraints on the training dataset to mitigate these biases, the authors observed a significant shift in the evaluation of detectors, yielding substantially improved robustness and generalization. The results demonstrate that the named biases significantly affect the detectors, leading to a misjudgment when evaluating. The authors emphasize the need for careful selection of training datasets to ensure that detectors do not inadvertently learn from undesirable variables.This paper investigates biases in generated image detection datasets, particularly focusing on JPEG compression and image size. The authors demonstrate that many datasets for AI-generated image detection contain biases related to JPEG compression and image size, which are subsequently used by detectors during inference. Using the GenImage dataset, they show that detectors indeed learn from these undesired factors. Removing these biases significantly enhances cross-generator performance, achieving state-of-the-art results on GenImage and increasing the average accuracy by more than 11 percentage points for ResNet50 and Swin-T detectors. Detectors become more robust against distortions due to now learning the actual task of detecting generation-specific artifacts. The GenImage dataset is one of the largest and most diverse datasets for generated image detection. It includes natural images from the ImageNet1k dataset and approximately an equal number of generated images from various generators. However, the evaluation of a detector based on the raw GenImage Benchmark is not reliable since JPEG compression and image size biases are used by detectors during training. The paper highlights that the use of JPEG compression and image size biases in training data can lead to detectors learning to distinguish between natural and generated images based on these factors, rather than the generation-specific artifacts. This can result in overestimation of detection performance and hinder the adaptability of detectors to real-world scenarios where such biases are absent. The authors conducted experiments to assess whether detectors trained on datasets containing such biases inadvertently acquire information from these undesirable variables. They found that removing these biases significantly improves the robustness and cross-generator performance of detectors. Specifically, they found that detectors trained on a constrained dataset with JPEG compression and size biases removed showed significantly improved performance compared to those trained on the raw GenImage dataset. The paper concludes that biases in generated image detection datasets can significantly affect the performance of detectors, leading to misjudgments in real-world scenarios. By imposing constraints on the training dataset to mitigate these biases, the authors observed a significant shift in the evaluation of detectors, yielding substantially improved robustness and generalization. The results demonstrate that the named biases significantly affect the detectors, leading to a misjudgment when evaluating. The authors emphasize the need for careful selection of training datasets to ensure that detectors do not inadvertently learn from undesirable variables.