Sora Detector
Sora clips are among the more identifiable generated videos, at least at first: OpenAI signs them with Content Credentials, and public exports carry a moving watermark. Both disappear the moment someone crops and re-encodes — which is exactly when frame analysis matters.
How Sora output is identified
Three layers, in descending order of convenience — and only the last one survives a determined re-upload.
Content Credentials
OpenAI embeds signed C2PA manifests in Sora output, marking it as AI-generated. Definitive while intact.
The moving watermark
Public Sora exports have carried a visible animated watermark. Croppable, and routinely cropped.
Re-encoding erases both
Download, crop, re-upload to a platform — manifest gone, watermark cut. Nothing above helps at that point.
Temporal drift
Across a longer shot, object detail and proportions subtly reconstruct themselves frame to frame. Real footage does not do this.
Physics that decays
Contact between objects, cloth, liquids and crowds hold up briefly, then stop obeying consistent rules.
Frame-level generation traces
Every extracted frame still carries the statistical fingerprints of generation — frequency profile, absent sensor noise, flat texture coherence.
How the analysis works
Upload the original file
Full resolution, straight from the source. Screenshots and re-saved copies destroy most of the evidence we measure.
12 independent methods run in parallel
Three neural networks look at the image as a whole. Nine forensic methods examine it as a signal: frequency (FFT) analysis, noise pattern, pixel-level statistics, EXIF metadata, Error Level Analysis, texture coherence, colour distribution, Benford's Law and edge detection.
You see the breakdown, not just a number
Every method reports separately, so you can tell whether the verdict rests on one weak signal or on nine that agree.
For still images rather than video, Start with our AI image detector — it covers any image type and runs the same twelve methods.
How much each method counts
Most detectors give you one number from one model. We run twelve independent analyzers and weight them — three neural networks carry the majority, but nine forensic methods examine the file as a signal and can overrule them. Here is the exact weighting we use in production:
| Method | Weight in final score |
|---|---|
| Neural network #2 | 25% |
| Neural network #1 | 15% |
| Neural network #3 | 15% |
| EXIF metadata | 10% |
| Frequency (FFT) | 6% |
| Benford's Law | 6% |
| Noise pattern | 5% |
| Error Level Analysis | 4% |
| Texture coherence | 4% |
| Edge detection | 4% |
| Pixel statistics | 3% |
| Colour distribution | 3% |
What people actually upload
These are our own numbers, not an industry estimate — every check run through FakeSec since launch. The split is closer than most people expect: roughly four in ten uploads really do turn out to be generated.
Where it matters
- Newsrooms — A clip presented as filmed footage.
- Platforms — Synthetic-media disclosure requirements.
- Advertisers — Verifying what was actually shot.
- Disputes — Video offered as a record of events.
Accuracy — and where it breaks down
On our internal benchmark of 240,000 images (120,000 AI-generated, 120,000 real) the combined score reaches up to 98% on unedited, full-resolution files. That number drops, sometimes sharply, in four situations — and you should know them before trusting any result:
- Heavy compression. Social platforms re-encode everything. A photo pulled from Instagram has already lost much of its noise signature.
- Screenshots. A screenshot is a new image of an image. Original metadata and compression history are gone.
- Stylised content. Flat colour and clean lines remove the texture and noise evidence forensic methods depend on.
- Post-processing. Added grain, filters, upscaling and manual retouching all mask generation traces.
No detector on the market is 100% accurate — ours included. Treat a high score as strong evidence, not as proof. When a decision matters, look at which of the 12 methods agreed.
Video is the hardest case in this whole category. Platform re-encoding destroys much of the forensic signal, so a clip pulled from social media gives a weaker answer than the original file. Upload the best-quality version you can obtain.
Frequently asked questions
Related tools
Check your first image now
Create a free account and get 3 checks. Full breakdown across all 12 methods, no card required.