Verified, then logged.
Sun 22 Nov 2026 · 00:13 KST / 15:13 UTC한국어 | ENSearch /

Breaking01 / AI

AI video detectors miss most fakes after reposting

Researchers identified changes in generators and content, alongside repost compression, as separate causes of weaker detection.

A researcher seen from behind checks an AI response on one monitor against numerous sources and scientific charts spread across three other monitors. AI-generated image

Researchers reported that no detector caught more than 15% of fake videos in an evaluation designed to reflect reposting conditions.

The study, published on arXiv on 13 Nov, tested 18 detectors. The results used a 1% false-positive rate, the share of real videos wrongly labelled fake. According to the abstract, existing detectors are typically evaluated on clean, curated footage.

The researchers introduced Reality Perception Gap (RPG), an evaluation set of 10,000 videos. They applied compression from reposting across platforms to generated videos and partially synthetic edits, pairing them with real footage retaining optical traces from cameras. The set also included crime and wartime footage, where fabricated videos can cause social harm.

The researchers used control and re-encoding experiments to separate the causes of failure. They found that changes in generators and content reduced detection performance before compression. Repost compression then erased the fine traces on which detectors that analyse visual artefacts depend.

The researchers proposed training under conditions that reflect actual deployment. They suggested combining fine visual traces with evidence from movement over time and the meaning of the content. Their findings also raised limits to detecting manipulation solely from clues left in a video.

The abstract did not disclose performance figures for the proposed combined approach. The researchers presented training under deployment conditions as a direction for improving detection. Their proposal brings together several kinds of evidence to address the separate failures identified in the experiments.