# AI video deep‑fake detectors perform poorly on repost‑compressed footage

> A new benchmark of 10,000 real‑world videos shows that 18 detectors catch fewer than 15% of fakes after social‑media compression.

Oossa · 2026-11-13 · https://oossa.com/en/ai-video-deep-fake-detectors-perform-poorly-on-repost-compressed-footage

Researchers released the Reality Perception Gap (RPG) benchmark, which adds repost compression and camera artifacts to 10,000 AI‑generated and edited videos. Testing 18 existing deep‑fake detectors on this set, none identified more than 15% of fake videos while keeping false alarms below 1%. The study says the drop comes from two sources: the shift in content and generator style before compression, and the loss of low‑level traces after compression. The authors suggest future detectors need training on such degraded data and should combine low‑level, temporal, and semantic clues.

## The facts

- 10,000 videos in the RPG benchmark
- Maximum 15% detection rate at 1% false‑positive across 18 detectors

## Why it matters

Law enforcement and platforms may miss most deep‑fake videos that have been re‑uploaded, limiting their ability to curb misinformation.

## Sources & references

1. [Reality Perception Gap: Are AI-Manipulated Video Detectors Ready for the Real World?](https://arxiv.org/abs/2611.14941) – arXiv cs.CV, 2026-11-13

Last updated: 2026-11-13