GenWorld: Towards Detecting AI-generated Real-world Simulation Videos

AI-generated videos pose a growing threat to information credibility: The paper demonstrates that advanced video generation technologies can now create highly realistic fake videos that replicate real-world scenarios, making it increasingly difficult for people to distinguish authentic content from AI-generated forgeries.

Existing detection tools have significant blind spots: Current AI-generated video detectors fail to identify high-quality videos created by state-of-the-art models like Cosmos, suggesting that many widely-deployed detection systems may not catch the most sophisticated fakes currently in circulation.

High-quality datasets are essential for building trustworthy detection systems: The researchers created GenWorld, a large-scale dataset featuring videos generated by multiple state-of-the-art models with diverse prompts (text, image, and video inputs), emphasizing that detector development requires realistic training data to be effective in real-world deployment.

Physical consistency offers a promising detection approach: The proposed SpannDetector method uses "multi-view consistency" as a detection criterion—essentially checking whether different angles and perspectives in a video maintain physical plausibility—offering a more explainable and interpretable way to identify AI-generated content compared to black-box detection methods.

Generalization across different generators matters for practical deployment: Because AI videos can be created using various generation models and input types, effective detectors must be tested and trained on diverse sources to avoid being fooled by new or unfamiliar video generation techniques.