Now that video generators can produce photorealistic "UFOs" in ten seconds, the first question is no longer "what is it?" but: did this video ever pass through a real lens? Good news: the physics of a real camera is hard to fake. And it can be measured.
A filmed video carries three physical signatures that generators imitate poorly. Here is how to check them — by eye first, by machine second.
On a real sensor, noise increases with brightness — a physical law (shot noise scales with the square root of the signal). Bright areas of a real video "crawl" more than dark ones. AI generators produce the opposite: flat, uniform, decorative grain — or no grain at all.
A hand-held video always shakes — a 1–8 Hz muscular tremor nobody can suppress voluntarily. Generated videos glide with unreal smoothness: the camera "floats" as if on an invisible rail, without the slightest vibration.
Many generators produce fewer real frames than advertised: they duplicate or interpolate frames to reach 30 fps. The result: an abnormal rate of near-identical frames, invisible during playback but measurable frame by frame.
| Clue | Real camera | Generated / heavily processed |
|---|---|---|
| Sensor noise | Grows with brightness (photon noise) | Flat, uniform or absent |
| Micro-shake | Present when hand-held (1–8 Hz) | Sweeping yet perfectly smooth motion |
| Frame cadence | Every frame is unique | Duplicated or interpolated frames |
None of these clues is proof. Heavy compression (WhatsApp, repeated re-shares) crushes sensor noise. Software stabilization smooths the shake. A tripod legitimately removes it. What should raise a flag is the accumulation of atypical clues combined with a missing original file — never a single clue in isolation.
And the other way around: a video that passes all three tests is not "proven authentic". The only real answer remains the original file, with its metadata, provided by the witness.
Eyeballing a noise-to-luminance correlation or a shake spectrum has its limits. VigiSky Sentinel measures it by machine: load the video, the app measures the three clues, tracks every moving object (trajectory, angular speed, light curve — it recognizes aircraft strobe lights by their frequency) and returns a cautious verdict, with its limits stated in plain words.
Download VigiSky Sentinel — free Windows · 100% local, your video never leaves your PC · authenticity clues + motion analysisNo. You can measure clues that make synthesis likely, but none is proof: aggressive compression produces similar signals. The only real answer is the original file with its metadata.
Trained classifiers fail with confidence and age with every new generation of models. Physical clues do not go out of fashion: photon noise is a law of the sensor, not a trend. Both approaches complement each other — ours has the advantage of being explainable line by line.
The shake clue becomes unusable — and an honest tool must say so rather than count it against the video. That is exactly what Sentinel reports: "camera perfectly still, clue unusable".
Ask for the original file (not a screenshot, not a re-share). Note the source and date. Run it through Sentinel. Facts first, interpretation second.