VideoVeritas: AI-Generated Video Detection via Perception Pretext Reinforcement Learning

The growing capability of video generation poses escalating security risks, making reliable detection increasingly essential. In this paper, we introduce VideoVeritas, a framework that integrates fine-grained perception and fact-based reasoning. We observe that while current multi-modal large language models (MLLMs) exhibit strong reasoning capacity, their granular perception ability remains limited. To mitigate this, we introduce Joint Preference Alignment and Perception Pretext Reinforcement Learning (PPRL). Specifically, rather than directly optimizing for detection task, we adopt general spatiotemporal grounding and self-supervised object counting in the RL stage, enhancing detection performance with simple perception pretext tasks. To facilitate robust evaluation, we further introduce MintVid, a light yet high-quality dataset containing 3K videos from 9 state-of-the-art generators, along with a real-world collected subset that has factual errors in content. Experimental results demonstrate that existing methods tend to bias towards either superficial reasoning or mechanical analysis, while VideoVeritas achieves more balanced performance across diverse benchmarks.

As AI-generated videos become increasingly realistic, distinguishing them from real footage is becoming a critical challenge. Existing tools for detecting fake videos tend to rely either on shallow pattern-matching or rigid rule-based analysis, making them unreliable across diverse scenarios. We introduce VideoVeritas, a system designed to detect AI-generated videos by combining two key abilities: carefully examining fine-grained visual details and reasoning about whether video content is factually plausible. The core insight is that while modern AI systems are good at reasoning, they often struggle to perceive subtle spatial and temporal details in videos. To address this, we train our system using simple auxiliary tasks — such as locating objects across frames and counting them — which sharpen its perceptual abilities without directly training it to detect fakes. To evaluate detection methods fairly, we also release MintVid, a curated dataset of 3,000 videos generated by 9 state-of-the-art AI systems, including real-world videos with factual inconsistencies. Experiments show that VideoVeritas achieves more balanced and robust detection performance compared to existing approaches.