A Vision-Language Dual-View Survey for factual fidelity verification
1 MBZUAI
2 Renmin University of China
3 Harvard University
† Corresponding author: Yuhan Liu · yuhan.liu@mbzuai.ac.ae
As video generators move from localized edits to fully synthetic, cinematic scenes, artifact-centric detection is no longer enough. This survey reframes AIGC-V detection as factual fidelity verification: whether the events, entities, and physical processes depicted in a video remain consistent with the real world.
Modern AI-generated video cannot be understood through frame artifacts alone. Detection must combine low-level intrinsic cues, spatiotemporal consistency, cross-modal consistency, and language-guided world-level reasoning.
The task is not only to label a clip as fake, but to identify which claim, event, entity, or temporal segment breaks consistency with reality.
The visual side covers intrinsic cues and spatiotemporal consistency inside the clip, while the language side covers cross-modal consistency and language-guided world-level reasoning.
Trustworthy systems should identify, localize, and explain with structured evidence instead of collapsing everything into a single score.
The survey organizes AI-generated video into three broad paradigms. Each one changes which cues are most useful, and therefore which detection strategies remain reliable.
The evidence profile shifts as generation moves from real-carrier manipulation to fully synthetic video synthesis.
Real footage is preserved while localized regions or attributes are re-rendered. The strongest evidence is often spatially concentrated and residual.
Speech drives the visual stream, making synchrony, speaker consistency, and speech-conditioned motion more decisive than generic image artifacts.
No authentic carrier remains. Detection shifts toward long-range coherence, world plausibility, provenance, and language-guided verification.
The taxonomy is organized by the dominant evidence pathway used for decision-making, not by a rigid task label. It moves from low-level perception to high-level cognition while keeping the boundary between layers operational.
Visual view: Layers 1-2. Language view: Layers 3-4. The boundary depends on what evidence the detector actually needs.
Focuses on the visual modality and emphasizes statistical differences between AIGC-V and real videos. The pathway extends from frame-level intrinsic cues to spatiotemporal consistency across frames, forming the perceptual perspective for factual-fidelity verification.
Tests whether low-level visual signals still follow the statistical regularities of real videos, including frequency fingerprints, local geometry, physiological traces, and residual synthesis artifacts.
Boundary L1-2: Layer 1 focuses on frame-level intrinsic distributional cues.Models frame-sequence relations to test whether motion, behavior, and physical transitions unfold plausibly over time as a real recorded video.
Boundary L1-2 / L2-3: Layer 2 moves beyond frames to explicitly model spatiotemporal connections across the frame sequence, but it still stays within the visual modality and evaluates within-visual coherence.Defines a grounded verification pathway in which audio, speech, text, and visual evidence are first checked for within-video consistency, then extended to factual and knowledge-level verification when external evidence is required.
Performs within-video multimodal verification across speech, text, and visuals, from lip-speech synchrony to voice-face identity coherence and speech-text agreement.
Boundary L2-3 / L3-4: Layer 3 goes beyond vision and relies on within-video cross-modal agreement, judging whether different modalities are mutually consistent in describing the same semantics.Moves beyond within-video agreement to verify implied claims against external facts, commonsense, physical rules, and explicit evidence chains.
Boundary L3-4: Layer 4 further requires external facts, commonsense, or physics for reasoning-based verification, even when the modalities appear mutually consistent.The landscape figure highlights the transition from artifact matching in traditional deepfake detection to evidence-based semantic verification enabled by vision-language models and agentic pipelines.
These cards summarize the broader migration pattern across layers, while the table below anchors that shift with the survey's complete-year counts through 2025.
| Year | L1 | L2 | L3 | L4 | Visual | Language | Language-View Share |
|---|---|---|---|---|---|---|---|
| 2020 | 5 | 7 | 1 | 0 | 12 | 1 | yearly / cumulative |
| 2021 | 5 | 6 | 4 | 0 | 11 | 4 | yearly / cumulative |
| 2022 | 4 | 3 | 2 | 0 | 7 | 2 | yearly / cumulative |
| 2023 | 3 | 3 | 4 | 0 | 6 | 4 | yearly / cumulative |
| 2024 | 9 | 5 | 4 | 3 | 14 | 7 | yearly / cumulative |
| 2025 | 0 | 18 | 15 | 15 | 18 | 30 | yearly / cumulative |
The mini-cards summarize the longer-term layer transition, and the year-wise trend table anchors it with the survey's complete-year summary through 2025; in-progress 2026 entries are intentionally kept outside this trend view.
Benchmark design should reveal why factual fidelity fails, where the failure occurs, and whether the system can ground its explanation with checkable evidence.
The deepest protocol tradition, emphasizing localized forensic residue, compression robustness, and cross-dataset transfer.
Smaller but tightly aligned with synchrony, speaker-content consistency, and timestamped localization.
Rapidly changing benchmarks targeting cross-generator transfer, semantic fabrication, plausibility, and world-level reasoning.
Adjacent diagnostics matter because trustworthy detection increasingly depends on physical plausibility and grounded explanations.
Browse the paper-first index by paradigm, detection layer, benchmark family, and adjacent diagnostic category.
Move cleanly across paradigms, detection layers, and benchmark families, then open the corresponding reference tables below.
Use Methods Overview for the taxonomy first, then switch to paradigms, layer pages, or benchmark families when you want a more specific slice of the literature.
Repository-level entry points and the complete flattened list.
Navigate by manipulation type and synthesis regime.
Follow the survey taxonomy from intrinsic cues to world-level reasoning.
Inspect evaluation families and adjacent diagnostic testbeds.
Select a route above to update the reading panel below.
Loading the selected library view.
A flat, paper-first index in the same style as the reference list. Detailed notes and extra metadata stay on the linked section pages.
The next phase is not just stronger clip-level discrimination. It requires diagnostic evaluation, claim-level supervision, unified explainable detection, and deployment protocols that preserve calibration when evidence is weak, conflicting, or incomplete.
Move beyond clip-level AUC or EER. Evaluation should reveal which factual proposition fails, where the violation occurs, and how brittle the detector becomes under shift, transfer, codec changes, and modern synthesis pipelines.
Benchmarks should move from clip labels to checkable claims with timestamped evidence, targeted stress tests, and continuously refreshed generator pools rather than static one-off test sets.
Trustworthy AIGC-V detection is a two-pathway problem: perceptual evidence plus fact-level verification. Cross-layer fusion should preserve low-level traces, localized mismatches, and tested claims as individually inspectable evidence objects.
Systems should identify, localize, and explain with calibrated uncertainty, provenance-aware cross-checks, and abstention when evidence remains incomplete or internally contradictory.
If you use this survey or its taxonomy in your work, please cite it as follows.
@misc{hou2026detecting,
title = {Detecting AI-Generated Video: A Vision-Language Dual-View Survey},
author = {Dylan Xinming Hou and Juntian Zhang and Xu Gu and Yichen Wu and
Nils Lukas and Gus Xia and Xiuying Chen and Yuhan Liu},
year = {2026},
note = {Survey manuscript},
howpublished = {\url{assets/paper.pdf}}
}