02
Video: what survives, and what stopped mattering
Blink rate was a real tell in 2018 and has not been one for years. What still works is consistency across time, and the reason clips are short.
6 min read
Generated and manipulated video has a different failure profile from still images, because it has to be right repeatedly rather than once.
That is the whole basis of what still works.
Time is the weak axis
A single frame can be flawless. Two hundred consecutive frames have to agree with each other about every detail in the scene, and agreement is expensive.
So the productive move is not to stare at the video. It is to freeze two frames a second apart and compare them.
Look at: earrings, moles, scars, stitching, buttons, background objects, the pattern on a wall. These should be identical. When they drift, migrate or change count, you have something.
Occlusion
Watch what happens when one thing passes in front of another. A hand crossing a face. A head turning past profile so that the far ear disappears and returns. Someone walking behind a pillar.
Occlusion forces the system to decide what is in front, hold the hidden thing in reserve, and restore it correctly. That sequence is hard, and it is where seams appear.
Mouths
The inside of a mouth during speech — teeth, tongue, the dark cavity behind them — remains difficult. Slow the footage and watch a long vowel.
Why the clip is short
This is a tell about production rather than pixels, and it is often the strongest one available.
Generated video is expensive and degrades over duration. Manipulated video is easiest over short, tightly framed, low-motion shots. So a claim resting on six seconds of a static head-and-shoulders, with no lead-in and no aftermath, is short for a reason.
Ask what happened in the minute before. Ask why there is only one angle of an event that would have had twenty phones pointed at it.
What to stop doing
Counting blinks. A genuine 2018 finding, comprehensively obsolete.
Judging by quality. Compression destroys artefacts in real and fake footage equally. Low resolution is not evidence of anything except low resolution.
Assuming live means real. Real-time face and voice replacement exists, runs on ordinary hardware, and is used in fraud today.
The thing that actually works
Corroboration. A real event that matters leaves more than one trace: another angle, a wire report, a timestamped post from someone unconnected, a building you can locate.
A fabricated event has exactly as many traces as its author made. Usually one.
Do not ask whether the video looks real. Ask what else exists.