What synthetic video does to believing anything

Two harms, and the larger one is not the obvious one

The obvious harm is fabrication: a video of someone doing something they did not do. It is real, it is documented, and its distribution is heavily skewed toward non-consensual sexual imagery of women and toward financial fraud.

The less obvious harm is the mirror of it. Once everyone knows video can be fabricated, any genuine recording can be dismissed as a fake. This is sometimes called the liar's dividend, and it does not require anybody to actually make a fake. It requires only that fakes are known to be possible.

This is a corrosion of a specific kind. Photographic and video evidence has, for about a century, functioned as a shared reference point that survived disagreement about everything else. What replaces it is not nothing, and it is slower, more institutional and less available to individuals: chains of custody, corroborating sources, and cryptographic provenance.

Why detection will not save it

The provenance module goes into the mechanisms. The short reason detection is the wrong hope:

• Detectors are trained on the outputs of known models and degrade sharply on models they have not seen, which includes every model released after the detector.
• Any lossy step — a re-encode, a screenshot, a platform upload — removes much of the statistical signal they rely on.
• The base rate problem is severe: apply a 95%-accurate detector to a population where 1 in 1000 videos is fabricated and most of your positives are wrong.
• And a detector that is public is a target that adversaries can optimise against.

The consequence: a detector's output is a weak signal, not a verdict, and using one to make a public accusation is how innocent people get harmed.

What actually works

Provenance at capture. Sign the recording when it is made, in the device, and carry the signature with the file. This proves origin rather than detecting fakery, and it is the only approach that improves as the generators improve. Its problem is adoption, not principle.

Corroboration. Multiple independent recordings of the same event, from angles that agree. This is how newsrooms have verified footage for years and it does not care how good the generators get.

Source over content. Who published this first, do they have a name, does the account have a history, does a reverse search find it earlier elsewhere. The free tools do this in seconds.

Out-of-band verification for anything consequential. Any instruction arriving by video or voice that moves money, grants access, or changes a payment detail gets confirmed on a different channel. Organisations that adopted this rule stopped losing money to this class of fraud immediately. It is the cheapest control in the whole field.

What to do as somebody who makes things

Two responsibilities, both small and both worth taking seriously.

Label your synthetic work. Not because you are ashamed of it, but because an unlabelled synthetic video that later circulates as real is a harm you contributed to. The final module covers what the law requires in several countries; the practice runs ahead of the law and should.

Do not build the demonstration. There is a persistent genre of "look how easy this is" content — a fabricated clip of a public figure, made to raise awareness. It raises awareness and it also adds a fabricated clip to the world, permanently, with your caption stripped by the second share. Show the workflow on yourself or a consenting colleague. The point lands identically.

There is one asymmetry worth naming, because it decides who bears the cost. Fabricating a convincing clip now takes minutes and a few pounds. Refuting one takes a newsroom, a forensics team and several days, and the refutation reaches a fraction of the audience the original did. That imbalance is not fixed by better tools on the defending side, because the effort ratio is the problem rather than the accuracy. It is the strongest practical argument for provenance at capture, which shifts the work to the moment of recording, where it is cheap, and away from the moment of dispute, where it is not.

The unsettled part, stated as such: nobody knows where this settles. Some argue that societies adapted to photographic manipulation before and will adapt again, with institutions and habits rather than technology doing the work. Others argue that the speed, volume and personalisation are different in kind. Both positions are held by serious people, the evidence is a few years old, and anyone confident in either direction is guessing. What can be said now is what to do about it, and the list above is not controversial.