A team spanning Peking University, Kuaishou's Kling group, HKUST (Guangzhou), Fudan University and Tsinghua University has introduced Verus-Insight-8B, a detector for AI-generated video that is designed to explain itself. The paper, submitted to arXiv on October 22, starts from a complaint about the current state of detection: existing artifact-based detectors can miss subtle artifacts, or reach the correct real/fake decision without faithfully identifying the evidence underneath it. A label, in other words, is not the same as a reason.

Verus-Insight splits the job into three parts. Inspection Planning decides what in a video is worth examining. Evidence Verification checks whether the inspected content actually shows a generation artifact. Evidence Aggregation combines the verified evidence into the final call. The model is built on Qwen3-VL-8B-Instruct and trained with human-anchored full-process supervision, including supervised fine-tuning and reinforcement learning. Alongside the model, the team built Verus-Bench, a benchmark of 192 videos with fine-grained human annotations of artifacts people can actually perceive. The training data was collected from the current generation of video models: Kling-2.5, Veo 3, HunyuanVideo-1.5, LTX-2.3 and Wan2.2.

Two scores, two different stories

The headline result is strong. On RewardData, an existing dataset, video-level balanced accuracy rises from 48.81 for the raw Qwen3-VL-8B backbone to 94.79. On Verus-Bench, the authors' own benchmark, it rises from 66.67 to 89.06. Read alone, those numbers suggest AI-video detection is close to solved.

The second metric in the same tables says otherwise. Verus-Insight achieves 94.79% balanced accuracy on RewardData, but its Finding F1 is 35.42, combining 33.66% precision and 37.36% recall. On Verus-Bench, it reaches 89.06% balanced accuracy with 15.97 Finding F1, 17.14% precision and 14.95% recall. Under the authors' semantic matching protocol, many annotated artifacts remain unmatched, while many predicted findings also fail to match the references.

Finding F1 needs defining, because it is easy to misread. The authors score findings by one-to-one semantic matching between predicted findings and human-annotated artifacts: a match requires the same target and the same specific abnormal phenomenon. Unmatched predicted findings count as false positives; unmatched reference artifacts count as false negatives. F1 combines precision and recall, so it is not a recovery rate. The recovery number is recall: under this protocol, the model matched about 37% of the annotated artifacts on RewardData and about 15% on Verus-Bench. The matching is semantic — the paper's appendix notes that the hits are semantic recovery events, not localisation successes, so the score does not itself verify where or when in the video an artifact occurs. The extraction and matching are also performed by an automated judge, Gemini 3.1 Pro, and the authors note that the stored automated matches do not constitute independent human agreement measurements.

The comparison the authors draw on RewardData makes the point sharper. In the Verus-Insight authors' RewardData comparison, VideoVeritas records 91.12% balanced accuracy and 8.47 Finding F1, with 4.79% precision and 36.26% recall: it recovers a similar share of the annotated artifacts, but roughly nineteen of every twenty findings it produces fail to match a human annotation. Any platform, newsroom or moderation team buying an artifact-based detection score should ask for precision, recall and the evaluation protocol alongside the accuracy figure, not just the headline number.

The release, checked

What was released, and on what terms, needs the same care. The authors have published inference code and linked model-weight and benchmark repositories. The code, in a public GitHub repository, is Apache-2.0 licensed, while the model weights and Verus-owned dataset materials carry non-commercial terms in their respective repositories; third-party benchmark media retain their original terms. The distributed Verus-Bench media cover 166 of the benchmark's 192 videos; the remaining 26 require separate acquisition. And availability is not fully confirmed: the linked Hugging Face model page returned an authorization error to an anonymous check on October 23, so whether the weights can be downloaded without an account or approval is unconfirmed. Availability of the linked downloads should be checked before describing the full release as publicly runnable.

What this changes in practice

For most readers the practical lesson is about how to read detector claims, not about this one model. A single accuracy figure, without an evidence metric beside it, can hide a system that has learned the statistical smell of generated video without being able to show a human reviewer anything checkable. When a detection result is used to take down a video, dispute a news clip or challenge an insurance or legal claim, the evidence trail is the part that has to survive scrutiny. A bare label does not.

There are limits to keep in view. Every figure here is the authors' own; no independent group has yet reproduced these results, and Verus-Bench was built by the same team that built the model, which is the usual arrangement and the usual caveat. RewardData is an existing dataset, but the results reported on it here are still author-run and author-scored. The training generators also date the work: a detector tuned on Kling-2.5, Veo 3 and Wan2.2 artifacts will face newer generators whose artifacts differ, and the paper does not claim otherwise.

Still, the direction is the right one for artifact-based detection. Artifact-based detectors that expose their inspection steps and findings give reviewers more to evaluate than a real/fake score alone. Whether those explanations are trustworthy still depends on independently testing their accuracy, coverage and consistency. That standard applies to detectors that produce explanations; watermark detectors and cryptographic provenance systems work differently, and this paper says nothing about them.

Sources

  1. [1] Zhang, Shi, Wang et al., “Verus-Insight: Towards Explainable Artifact Detection in AI-Generated Videos”, arXiv:2610.30866 (submitted October 22, 2026) — full text, Tables 2–3 and evaluation appendix A.4 read first-hand for this articleRead source
  2. [2] Verus-Insight code repository (zhangzhuoran997997/Verus-Insight), GitHub — README, including the reported-results table, the benchmark data notes (media for 166 of 192 samples) and the “Development and license” section; read October 23, 2026Read source
  3. [3] Hugging Face paper page for arXiv:2610.30866 (published October 22, submitted to HF October 23, 2026)Read source
  4. [4] AI Weekly, “Peking U, Kling open Verus-Insight-8B for AI-video artifacts” (October 23, 2026) — release reporting; every figure in this article was verified against the paper itselfRead source