AI-Generated Video Detection at Native Scale: 140K Videos, 15 Generators, New State-of-the-Art (ICLR 2026)

A new ICLR 2026 paper introduces the most comprehensive AI-generated video detection system to date, processing 140K+ videos from 15 generators without the information-destroying preprocessing that plagues existing methods.

The Problem with Current Detection

Existing AI video detectors suffer from a critical flaw:

• They resize and crop videos to fixed resolutions
• This destroys high-frequency forgery traces — the very signals needed for detection
• Preprocessing causes spatial distortion and significant information loss
• Training datasets are outdated and fail to capture modern generators

It's like examining a forged painting through a blurry lens — you're destroying the evidence you need to see.

The Solution: Native-Scale Processing

The researchers built a detection framework on Qwen2.5-VL Vision Transformer that operates at:

• Variable spatial resolutions — No fixed resizing
• Variable temporal durations — No fixed cropping
• Native scale — Preserving all high-frequency artifacts

The Dataset

Key Innovations

1. Preserving forgery artifacts — Working at native scale keeps subtle detection signals intact
2. Spatiotemporal inconsistencies — Captures timing and spatial errors that generators make
3. Comprehensive benchmark — Most extensive evaluation of AI video detection to date
4. Modern generators — Includes latest state-of-the-art video synthesis models

ICLR 2026 Camera Ready

The paper has been accepted and is in camera-ready form, indicating peer review validation.

Why It Matters

• Misinformation defense — Essential for combating deepfakes and synthetic media
• Content platforms — Social media companies need reliable AI video detection
• Legal evidence — Forensic tools for verifying media authenticity
• Arms race — Detection must keep pace with ever-improving generation models