Skip to content

Detecting AI-Generated Video: A Vision-Language Dual-View Survey

🕒 Published (v1): 2026-07-12 14:25 UTC · Source: Arxiv · Venue: ACL 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This survey reframes AI-generated video (AIGC-V) detection as Factual Fidelity Verification — asking whether a video's implied fact-level propositions are consistent with the real world — rather than binary real/fake classification. It proposes a Vision-Language Dual-View taxonomy organizing 221 works into a four-layer hierarchy from low-level artifact detection to VLM-powered world-level reasoning. The central thesis is that modern generative video quality necessitates a paradigm shift from perceptual artifact inspection to evidence-based semantic verification.

Problem

Traditional deepfake detectors rely on low-level visual artifacts (blending boundaries, frequency fingerprints) that are increasingly imperceptible in outputs from modern diffusion/transformer-based generators (Sora 2, Veo 3, Seedance 2.0). Existing surveys are either constrained to the pre-diffusion deepfake era or treat video as a sequence of independent frames, failing to account for methods that use language and world knowledge for semantic verification. No unified taxonomy bridges visual forensics and multimodal reasoning for AIGC-V.

Method

The paper proposes a Vision-Language Dual-View taxonomy with two complementary pathways and four layers:

  • Visual View (Layers 1–2): operates on visual modality alone
  • Layer 1 — Intrinsic Cue Analysis: frame-level distributional evidence (pixel/geometric artifacts, physiological signals such as rPPG, representation distribution discrepancy)
  • Layer 2 — Spatiotemporal Consistency: explicit inter-frame modeling of motion, temporal coherence, physical plausibility
  • Language View (Layers 3–4): introduces cross-modal and external knowledge
  • Layer 3 — Cross-Modal Consistency: audio-visual synchrony, text–visual semantic consistency (within-video, no external knowledge required)
  • Layer 4 — Language-Guided World-Level Reasoning: VLM-driven verification against external facts, commonsense, and physical laws; includes tool-augmented agentic pipelines and post-training for explainable reasoning (SFT + RL)

Detection is formally defined as a mapping \(D: \mathcal{X} \rightarrow [0,1] \times \mathcal{E}\), where the output couples a suspicion score \(s(x)\) with a structured evidence set \(e(x) = \{\epsilon_k\}_{k=1}^K\); each evidence item \(\epsilon_k = (g_k, c_k, \pi_k, r_k)\) encodes a grounding, confidence, claim-check statement, and provenance record. The ideal detection target is \(y(x) = \mathbf{1}[\Pi(x) \not\subseteq \mathcal{P}_\text{world}]\), where \(\Pi(x)\) is the set of fact-level propositions implied by the video.

Three AIGC-V paradigms are also taxonomized: Local Manipulation (face swap, attribute editing), Audio-Visual Editing (talking heads, lip-sync), and Generative Video Synthesis (text-to-video, end-to-end).

Key Contributions

  • Reformulation of AIGC-V detection as Factual Fidelity Verification with a formal grounded-evidence output specification
  • Vision-Language Dual-View taxonomy: a four-layer, evidence-boundary-criteria-based classification of 221 works spanning 2018–2026
  • Systematic coverage of AIGC-V generation paradigms (Local Manipulation, Audio-Visual Editing, Generative Video Synthesis) and their distinct forensic implications
  • Characterization of the shift from artifact matching to VLM-based semantic and world-level reasoning as a first-order trend in the field
  • Unified review of evaluation metrics and benchmarks aligned with the dual-view framing

Results

This is a survey paper; no original experimental results are reported. The survey covers 221 works through March 2026. Illustrative detection pipeline outputs depicted in Figure 1 show metrics such as Frame AUC = 0.98, Video AUC = 0.95, Temporal F1 = 0.91, and Text–Video Similarity = 0.34, but these are example detector outputs used to illustrate the framework, not aggregated survey findings.

Limitations

  • Survey coverage is through March 2026; the field evolves rapidly and methods post-cutoff are not included
  • Layer 1 (intrinsic cue) methods degrade severely under compression, transcoding, and domain shift; physiological cues additionally require stable facial visibility and are brittle under occlusion and re-encoding
  • Layer 1 and 2 methods have weakening cross-generator transfer as GVS model diversity grows
  • The "language view" framing of audio (non-linguistic and linguistic cues) may blur boundaries with purely acoustic methods not grounded in language
  • The paper text provided is truncated at Layer 2, so the survey's stated limitations for Layers 3–4, benchmark gaps, and future directions cannot be fully characterized from the available text

Relevance to Vision-Language Models

This survey directly positions VLMs as the enabling technology for the frontier of AIGC-V detection (Layer 4), arguing that their capacity for world-knowledge grounding, cross-modal reasoning, and agentic tool use is necessary to detect generation artifacts that are no longer visually distinguishable. It catalogs how VLMs are being adapted — through language-calibrated forgery representations, RAG-augmented evidence retrieval, and RL/SFT post-training — specifically for forensic tasks, making it a direct reference for researchers studying VLM reasoning capabilities and their real-world application. The dual-view taxonomy also provides a structured lens for understanding where VLMs add value over purely visual models, which is relevant to grounding, consistency verification, and claim-checking research threads within the VLM community.