Real Data Lies: Unveiling and Closing the Quality Shortcut in Generalizable AI-Generated Video Detection

Recent advances in video generation have enabled highly realistic synthetic content, raising concerns about the integrity of digital media and motivating the development of benchmarks and detection methods for generated videos. Prior works have largely prioritized bolstering model generalization against unseen generators. However, we uncover a neglected factor: the quality distribution of real videos plays a pivotal role. Current training protocols suffer from a clear quality bias between real and fake data, prone to shortcut learning. Compounded by testing on similar real data distributions, this creates an illusion of generalization. In reality, these models fail to generalize when exposed to real data with significantly different quality profiles. To address this, we propose training with quality-matched real and fake data to mitigate bias. Building on this, we introduce a data expansion strategy that broadens the training set to comprehensively cover the full quality spectrum. This approach enables the model to learn quality-agnostic features for detection, thereby achieving generalization across real data of varying qualities and enhancing real-world applicability. Extensive experiments demonstrate that our method scales well across diverse backbones, consistently enhancing the generalization capability of existing models.

AI-generated videos are becoming incredibly realistic, making it hard to tell what is real and what is fake. Researchers have built detection tools that claim to spot these fake videos even when faced with new types of AI generators they have never seen before. However, our study reveals a hidden flaw in how these tools are trained and tested. We found that current detectors often cheat by relying on video quality. During training, the real videos used are typically lower in quality than the AI-generated ones. As a result, the detectors learn a simple but wrong rule: "high quality means fake." This works when testing on similar low-quality real videos, but fails dramatically when encountering high-quality real footage, such as professional or ultra-HD videos. To fix this, we propose matching the quality of real and fake training videos so the model cannot use quality as a shortcut. We also expand training data to cover the full range of video quality found in the real world, from blurry clips to crystal-clear footage. This forces the detector to learn genuine signs of forgery rather than simple quality cues. Our experiments show this approach significantly improves reliability across different video qualities. We also highlight that current benchmark tests are incomplete because they only use narrow quality ranges, and we suggest a more comprehensive evaluation protocol to better reflect real-world conditions.