Dylan Xinming Hou, Juntian Zhang, Xu Gu, Yichen Wu, Nils Lukas, Gus Xia, Xiuying Chen, Yuhan Liu
5 min
The evolving realism of AI-generated Videos (AIGC-V) is rapidly rendering traditional artifact-centric detection insufficient, necessitating a paradigm shift from low-level inspection to high-level semantic verification. This paper presents a comprehensive survey of AIGC-V detection, reframing the task as Factual Fidelity Verification, which asks whether the events, entities, and physical processes depicted in a video are consistent with real-world facts. To systematize this rapidly evolving field, we propose a Vision-Language Dual-View taxonomy that organizes existing methods into a hierarchical, four-layer landscape, spanning intrinsic cue analysis, spatiotemporal consistency modeling, cross-modal consistency reasoning, and language-guided world-level reasoning. This dual-view framing highlights a fundamental transition from artifact matching in traditional deepfake detection to evidence-based semantic verification enabled by vision-language models and agentic reasoning pipelines. Based on a systematic review of 221 works, we synthesize AIGC-V generation paradigms, survey the landscape of detection methods, and review evaluation metrics and benchmarks in line with proposed views. Finally, we discuss current challenges and identify promising directions toward robust, explainable, and trustworthy detection.
As generative models like Sora and Veo achieve cinematic realism, traditional detection methods—which rely on identifying low-level visual artifacts like blending boundaries or pixel inconsistencies—are becoming increasingly obsolete. The authors argue that the field must transition from "artifact-centric" detection to "Factual Fidelity Verification." This new paradigm treats video content as a series of propositions about entities, events, and physical processes, asking whether these propositions are consistent with real-world facts rather than simply checking if a video was synthesized by an AI.
To organize the rapidly expanding landscape of detection research, the authors introduce a Vision-Language Dual-View taxonomy. This framework classifies methods based on the evidence they prioritize:
This dual-view approach is further structured into a four-layer hierarchy that tracks the progression of detection technology:
This survey provides a structured roadmap for researchers to move beyond the "cat-and-mouse" game of artifact detection. By emphasizing explainable, evidence-based verification, the authors highlight a path toward more trustworthy systems that can provide human-readable justifications for their decisions, which is critical for combating misinformation in an era of high-fidelity synthetic media.
Sam: And the fourth layer is where it gets really interesting, right? That's the one that goes beyond the video entirely?
Alex: Exactly. The fourth layer sends the system out into the world, so to speak. It uses external knowledge — databases, encyclopedias, verified records — to check whether the video contradicts established facts. If a video shows a building in a specific city, the system can check whether that building actually exists there. If a billboard appears in the background, it can verify whether that brand name is real. The system is essentially acting as a fact-checker.
Sam: That's a meaningful shift. But I can imagine it being quite slow. Searching external databases for every claim in a video sounds computationally expensive.
Alex: It does add significant complexity. The paper describes this fourth layer as using what it calls "agentic reasoning" — meaning the system doesn't just look up one fact and stop. It behaves more like a researcher: it forms a hypothesis about what might be wrong, searches for relevant evidence, evaluates what it finds, and then updates its conclusion. It's an iterative process.
Sam: So the output isn't just a "fake" or "real" label. It's more like a structured argument — here are the claims we checked, here's the evidence we found, here's our conclusion.
Alex: That's the key advantage. Rather than a black-box verdict, the system produces what the paper calls evidence grounding — a human-readable trail showing exactly which parts of the video were verified, and where the inconsistencies appeared. That makes the process auditable, not just automated.
Sam: Though I imagine that also means the system is only as good as its knowledge sources. If the external database is outdated or incomplete, the whole fourth layer could give you a wrong answer.
Alex: That's a genuine limitation the paper acknowledges. The system's reliability is tied to the quality and freshness of its knowledge, and like any AI reasoning system, it can occasionally misinterpret or misapply what it finds. The goal isn't to replace human judgment — it's to give human reviewers a structured, evidence-backed starting point.
Sam: So it's less of a perfect shield and more of a transparent process. You can see where the evidence is strong and where it's uncertain.
Alex: Precisely. And that transparency matters more as the stakes get higher. A system that shows its reasoning is far more useful in a legal or journalistic context than one that simply outputs a confidence score.
Sam: Do you think this kind of approach will eventually become standard — the way we routinely verify media before trusting it?
Alex: The paper suggests it's a necessary direction. As AI-generated video becomes increasingly difficult to distinguish from real footage on visual grounds alone, verifying factual consistency may become as routine as checking the source of a news article. It won't be a complete solution on its own, but it represents a meaningful evolution in how we think about digital trust.
Sam: It's a sober reminder that the challenge isn't just technical — it's about building systems we can actually reason with, not just systems that give us an answer. Thanks for walking me through this, Alex.
Alex: Thank you, Sam. That's our look at factual fidelity verification in AI-generated video. Thanks for listening to ResearchPod.