news

A Kling-affiliated AI-video detector hits 89.06 balanced accuracy and 15.97 on finding the artifacts

Verus-Insight-8B and the Verus-Bench dataset are open, and the paper's own tables show one rival detector calling real-versus-fake correctly 91.12% of the time while recovering almost none of the flaws humans marked.

By Priya Shenoy · 24 October 2026

A detector that labels an AI-generated clip correctly is not necessarily looking at anything real when it does so. That is the claim at the center of Verus-Insight, an 8B vision-language model and companion benchmark posted on October 22, and the paper backs it with a number: a competing specialist detector, VideoVeritas, records 91.12 balanced accuracy on real-versus-fake calls alongside a Finding F1 of 8.47, meaning it recovers close to none of the specific artifacts human annotators marked in the same videos.

The work comes from a team whose affiliations span Peking University, Kuaishou's Kling video group, HKUST (Guangzhou), Fudan and Tsinghua. Weights, code, the benchmark and a project homepage are public, with the code under Apache 2.0 and the model runnable through Transformers or vLLM.

Detection is split into planning, verification and aggregation

Rather than treating detection as one classification step, the authors break it into three capabilities: Inspection Planning, which decides what to examine, Evidence Verification, which checks whether the inspected region actually shows a generation artifact, and Evidence Aggregation, which combines verified evidence into the final source call. The published pipeline runs context, rubric, audit, aggregation in sequence. Training starts from Qwen3-VL-8B-Instruct, with a first stage that jointly trains region tracking and rubric planning at a 1:1 sampling ratio before full-process supervised fine-tuning and reinforcement learning.

The reported gains over that backbone are large on the verdict metric and much smaller on the evidence metric. Video-level balanced accuracy moves from 48.81 to 94.79 on RewardData and from 66.67 to 89.06 on Verus-Bench. Finding F1 moves from 3.38 to 35.42 on RewardData and from 7.56 to 15.97 on Verus-Bench. The second pair is the one worth sitting with: on the benchmark the authors built themselves, their own model recovers under 16% of human-annotated artifacts by that measure while being right about provenance roughly nine times in ten.

Verus-Bench draws its real footage from NExT-QA, OneThinker-STVG, DiDeMo and an internet collection. The published dataset snapshot lists 192 videos split evenly at 96 real and 96 generated, with 320 findings across 17 artifact categories, and media distributed for 166 of the 192 samples. That is a small benchmark by the standards of the leaderboards this site tracks, and the authors present it as a fine-grained evaluation layer on top of source prediction rather than a replacement for larger detection corpora.

Why a generation lab is funding its own artifact detector

Kling's presence on the author list is the commercially interesting part. A video model vendor gets more practical value from a detector that localizes which artifact occurred in which region at which timestamp than from a binary real-or-fake classifier, because localized failures feed back into training and into reward modeling. Our own scoring of Kling 4.0 breaks quality down by failure type for the same reason, since hands, liquids and on-screen text fail in different ways and a single aggregate score hides that.

The policy context makes the evidence gap matter beyond research. OpenAI and Meta told a US evidence-rules panel earlier this year that no method reliably detects AI-generated video. Verus-Insight does not contradict that, and arguably sharpens it: a high balanced accuracy number, the figure most detection vendors quote, can coexist with a near-zero ability to say what in the frame is wrong. Anyone treating a detector score as courtroom-grade or platform-grade evidence should be asking for the second column as well as the first.

Sources: arXiv, Hugging Face, Verus-Insight on GitHub.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.