Verus-Insight: Towards Explainable Artifact Detection in AI-Generated Videos

The paper frames AI-generated video detection as a structured process of Inspection Planning, Evidence Verification, and Evidence Aggregation. It introduces Verus-Insight, trained with human-anchored full-process supervision and progressive capability training, and Verus-Bench, a benchmark with fine-grained human annotations of perceptible artifacts. The authors report strong performance on fine-grained artifact analysis and video-level detection, with gains in recovering human-perceived artifacts.

Key takeaways

• Artifact detection is decomposed into inspection planning, evidence verification, and evidence aggregation.
• Verus-Insight is trained with human-anchored full-process supervision and progressive capability training.
• Verus-Bench provides fine-grained human annotations of perceptible artifacts for evaluating artifact detection.
• Reported gains include recovering human-perceived artifacts without increasing spurious artifact claims.

Why it matters

Detectors that only predict a source label may reach correct decisions without faithful evidence. Explainable, evidence-grounded detection could make AI-generated video judgments more auditable as synthetic video becomes more realistic.

Read with care

Source claims not independently checked; the reported results come from the authors' own experiments and benchmarks.