The paper introduces the first million-scale dataset for AI-generated video detection, named GenVideo, which includes over one million AI-generated and real videos. The dataset features a rich diversity of content and methodologies, covering a broad spectrum of video categories and generation techniques. To evaluate the performance of detectors, two tasks are proposed: cross-generator video classification and degraded video classification. The cross-generator task assesses the generalizability of detectors on unseen generators, while the degraded video task evaluates their robustness against video quality degradation. Additionally, the paper introduces DeMamba, a plug-and-play module designed to enhance detectors by identifying AI-generated videos through the analysis of spatial-temporal inconsistencies. Extensive experiments demonstrate that DeMamba significantly improves the generalizability and robustness of detectors on GenVideo compared to existing methods. The authors believe that the GenVideo dataset and DeMamba module will significantly advance the field of AI-generated video detection.The paper introduces the first million-scale dataset for AI-generated video detection, named GenVideo, which includes over one million AI-generated and real videos. The dataset features a rich diversity of content and methodologies, covering a broad spectrum of video categories and generation techniques. To evaluate the performance of detectors, two tasks are proposed: cross-generator video classification and degraded video classification. The cross-generator task assesses the generalizability of detectors on unseen generators, while the degraded video task evaluates their robustness against video quality degradation. Additionally, the paper introduces DeMamba, a plug-and-play module designed to enhance detectors by identifying AI-generated videos through the analysis of spatial-temporal inconsistencies. Extensive experiments demonstrate that DeMamba significantly improves the generalizability and robustness of detectors on GenVideo compared to existing methods. The authors believe that the GenVideo dataset and DeMamba module will significantly advance the field of AI-generated video detection.