Kling and Peking University publish a detector that shows its work

Most AI-generated video detectors give you one word: real or fake. Verus-Insight, an 8B-parameter model posted to arXiv on October 22 by a joint Peking University and Kling team, is built to give you the reasons as well. It labels a video and then points to the specific artifacts it inspected to get there.

That distinction matters to anyone who has to argue about a clip. A platform enforcing a provenance rule, a client disputing a charge, or a creator defending a real upload all need more than a verdict. They need to know which frame, which region and which visible flaw drove the decision. A detector that lands on "AI-generated" for the wrong reason, or no stated reason, cannot support any of that.

What Verus-Insight actually does

The paper, titled "Verus-Insight: Towards Explainable Artifact Detection in AI-Generated Videos", breaks detection into three named steps. Inspection Planning decides what to look at in a given video. Evidence Verification checks whether the inspected content really shows a generation artifact, with spatial and temporal support where available. Evidence Aggregation then combines the verified evidence into the final real-or-fake call.

The authors trained the model on what they call human-anchored full-process supervision, built from artifact annotations that human reviewers made frame by frame, and on progressive capability training that runs through tracking and rubric warm-up, full-process fine-tuning, and reinforcement learning. The base model is Qwen3-VL-8B-Instruct.

The team spans Peking University, Kuaishou's Kling group, HKUST-Guangzhou, Fudan University and Tsinghua University. The paper lists 14 authors, including Zhuoran Zhang, Yang Shi and Qixun Wang, and Pengfei Wan and Yuanxing Zhang of the Kling team.

What the numbers show

On video-level detection, the paper reports that Verus-Insight beats its own backbone: balanced accuracy on RewardData rises from 48.81 to 94.79, and on the team's Verus-Bench from 66.67 to 89.06. On Artifact-Bench's T1 track it improves from 50.80 to 66.60.

The artifact-recovery side is where the honest picture sits. Finding F1, which measures how many human-marked artifacts the model actually recovers, climbs from 7.56 to 15.97 on Verus-Bench and from 3.38 to 35.42 on RewardData. Those are large gains and low absolute scores at once. A Finding F1 of 35.42 means the model still misses most of what annotators marked.

The paper makes that gap its central point. It reports that VideoVeritas, a specialized detector, reaches 91.12 balanced accuracy with a Finding F1 of only 8.47. A near-perfect real-or-fake score can sit on almost no artifact evidence. Verus-Insight is designed so the score and the evidence move together, and the authors report it converts more of its inspection targets into supported findings without adding spurious ones.

Why explainable detection changes the argument

For provenance and takedowns, the value is in what a reviewer can check. A detector that names the artifact, its frame range and its location gives a platform something to log and a creator something to contest. A bare label gives neither side anything to work with.

The paper also introduces Verus-Bench to test that skill directly. The benchmark holds 192 videos, 96 real and 96 AI-generated, with artifact annotations that record an observation, its time span and frame-wise bounding boxes. The generated clips were collected from recent models including Kling-2.5, Veo 3, HunyuanVideo-1.5, LTX-2.3 and Wan2.2. The paper counts 326 findings across 18 categories; the project page and the GitHub release say 320 findings across 17 categories, a small discrepancy between the paper and its own release notes.

An AI Weekly alert first flagged the release. The wider field is moving the same way: detectors are being graded not only on the verdict but on whether they can show their work, and Verus-Insight is one of the first public attempts to make that the target rather than an afterthought.

The inference code is public on GitHub under Apache-2.0, but the Hugging Face pages the paper links for the weights and the benchmark did not resolve publicly, so running Verus-Insight today means starting from the repository and waiting for the weights to appear.