This paper addresses the challenge of detecting AI-generated images, a critical issue in the rapidly evolving field of generative models. The authors introduce the Chameleon dataset, which consists of AI-generated images that are genuinely challenging for human perception. They evaluate nine off-the-shelf AI-generated image detectors on this dataset and find that all models struggle to distinguish AI-generated images from real ones, often classifying them as real. To improve the detection performance, the authors propose AIDE (AI-generated Image DETector with Hybrid Features). AIDE leverages multiple experts to extract both low-level visual artifacts and high-level semantic information. Specifically, it uses CLIP to compute visual embeddings and selects high-frequency and low-frequency patches to capture noise patterns. The model then fuses these features in the channel dimension for final prediction. AIDE is evaluated on two popular benchmarks, AIGCDetectBenchmark and GenImage, achieving improvements of +3.5% and +4.6% over state-of-the-art methods, respectively. On the Chameleon dataset, AIDE also shows promising results, despite the ongoing challenges in detecting AI-generated images. The paper concludes by highlighting the limitations of existing approaches and the need for more realistic evaluation benchmarks. The dataset, codes, and pre-trained models are available at <https://github.com/shilinyan99/AIDE>.This paper addresses the challenge of detecting AI-generated images, a critical issue in the rapidly evolving field of generative models. The authors introduce the Chameleon dataset, which consists of AI-generated images that are genuinely challenging for human perception. They evaluate nine off-the-shelf AI-generated image detectors on this dataset and find that all models struggle to distinguish AI-generated images from real ones, often classifying them as real. To improve the detection performance, the authors propose AIDE (AI-generated Image DETector with Hybrid Features). AIDE leverages multiple experts to extract both low-level visual artifacts and high-level semantic information. Specifically, it uses CLIP to compute visual embeddings and selects high-frequency and low-frequency patches to capture noise patterns. The model then fuses these features in the channel dimension for final prediction. AIDE is evaluated on two popular benchmarks, AIGCDetectBenchmark and GenImage, achieving improvements of +3.5% and +4.6% over state-of-the-art methods, respectively. On the Chameleon dataset, AIDE also shows promising results, despite the ongoing challenges in detecting AI-generated images. The paper concludes by highlighting the limitations of existing approaches and the need for more realistic evaluation benchmarks. The dataset, codes, and pre-trained models are available at <https://github.com/shilinyan99/AIDE>.