This paper introduces GenVideo, a large-scale dataset for AI-generated video detection, containing over one million AI-generated and real videos. The dataset is designed to evaluate the generalization and robustness of video detectors in real-world scenarios. It features a diverse range of video content and generation methods, covering various video categories and generation techniques. The authors propose two evaluation tasks: cross-generator video classification, which assesses the generalizability of detectors across different generators, and degraded video classification, which evaluates the robustness of detectors to handle videos with quality degradation. They also introduce a plug-and-play module called Detail Mamba (DeMamba), which enhances detectors by analyzing inconsistencies in spatial and temporal dimensions to identify AI-generated videos. Extensive experiments show that DeMamba outperforms existing detectors in terms of generalization and robustness on GenVideo. The authors believe that GenVideo and DeMamba will significantly advance the field of AI-generated video detection. The dataset and code are available at https://github.com/chenhaoxing/DeMamba.This paper introduces GenVideo, a large-scale dataset for AI-generated video detection, containing over one million AI-generated and real videos. The dataset is designed to evaluate the generalization and robustness of video detectors in real-world scenarios. It features a diverse range of video content and generation methods, covering various video categories and generation techniques. The authors propose two evaluation tasks: cross-generator video classification, which assesses the generalizability of detectors across different generators, and degraded video classification, which evaluates the robustness of detectors to handle videos with quality degradation. They also introduce a plug-and-play module called Detail Mamba (DeMamba), which enhances detectors by analyzing inconsistencies in spatial and temporal dimensions to identify AI-generated videos. Extensive experiments show that DeMamba outperforms existing detectors in terms of generalization and robustness on GenVideo. The authors believe that GenVideo and DeMamba will significantly advance the field of AI-generated video detection. The dataset and code are available at https://github.com/chenhaoxing/DeMamba.