AI or filmed?
The evidence hierarchy.

Why a verdict is hard

A video is a stack of evidence, and each layer can be destroyed independently. A camera writes sensor noise, a codec quantizes it, a platform re-encodes it, a downloader strips the container metadata, and a screen recording adds a second generation of artifacts. By the time a clip reaches a browser, the strongest evidence may be gone.

That is why “AI detector” is a misleading product category. The honest version is a probability estimate with a list of signals and a confidence level. The browser workbench deliberately stops before the verdict.

The evidence hierarchy

Layer	Examples	Strength
Container / provenance	AIGC implicit labels, C2PA manifests, device make/model, encoder signature	Strong when present; often stripped
Frame cadence	24 / 25 / 30 / 60 fps fingerprints, repeated-frame patterns, interpolation cadence	Medium-high on original files
Sensor and texture	Temporal noise, grain structure, flat 8×8 blocks, color count	Medium; compression can mimic or erase
Optics and physics	Depth of field, motion blur, reflections, shadows, contact points	Medium; a vision model or a careful eye is needed
Semantics	Hands, text, teeth, object permanence, physical continuity	Weak alone; strong in clusters
Platform context	Watermarks, upload patterns, account behavior	Circumstantial

The hierarchy matters because a visual oddity is often the weakest evidence and the easiest to over-read. A garbled sign is a signal; a garbled sign plus a 24 fps cadence plus no sensor noise plus a flat background is a pattern.

What the browser sees

The workbench measures cuts, luma, contrast, palette, edge density, flat-block ratio, estimated motion and an audio beat grid. Those are useful signals: a fully static frame with a flat texture is not live footage in the usual sense; a 24 fps cadence with no sensor noise is worth a second look. But the browser does not read the container, the AIGC tag or the C2PA manifest, and it does not understand the subject.

The report therefore says “signals”, not “verdict”. It will tell you that a shot is static, low key and texturally flat. It will not tell you that it is AI.

What the CLI adds

The open-source rea-video CLI runs ffmpeg locally and adds the layers the browser cannot reach: container and device metadata, AIGC labels, C2PA manifests, OCR text tracks, separated voice/music/effects, and an optional vision-model review. It fuses the deterministic evidence and the visual review into a probability per shot, with the evidence and a confidence score in the report.

The project publishes accuracy figures from its own evaluation sets — 98.3% on its tuning set, 99.6% / 98.9% on its holdouts, 88.6% on re-encoded WeChat videos with offline tools alone. Those are the project’s self-reported results on its own samples, not this site’s. Read the methodology in its eval before quoting a number.

How to say it

Label uncertainty. “This shot has AI-like signals” is defensible; “this creator faked it” is a different claim, and it needs a much higher bar. A false accusation travels faster than a correction. If the decision is legal, financial or reputational, use the CLI, keep the original file, and have a human review the evidence.

The workbench exists to make the first pass fast and honest, not to turn a heuristic into an accusation.