Tracking: video and audio detection has no deterministic technique yet

[mgiovani]

Status

Tracking only. No deterministic video or audio technique reaches useful precision today; this issue records the evidence so the question is not re-opened without new data.

Evidence (2026-09)

• Cakiroglu et al. 2026, detection from H.264 motion fields in the compressed bitstream, the only model-free method found: AUC 0.64. https://arxiv.org/abs/2607.19476
• RobustSora 2025: current video detectors mostly key on visible generator watermarks and drop 9–14 points when those are removed. https://arxiv.org/abs/2512.10248
• Michels et al. 2026, dataset biases and shortcut learning in motion-based detectors: https://arxiv.org/abs/2607.00948
• AudioSeal https://arxiv.org/abs/2401.17264 (https://github.com/facebookresearch/audioseal) and WavMark https://arxiv.org/abs/2308.12770: open detectors, but trained networks, and only for audio the generator chose to mark.
• Benchmarks: AIGVDBench https://arxiv.org/abs/2601.11035, GenVidBench https://arxiv.org/abs/2501.11340

Only deterministic candidate

Container metadata: MP4 ©too / encoder atoms and WebM WritingApp naming a generator. No paper reports its precision. Revisit when one does, or when a generator ships a public, key-free watermark detector.