I think it’s already irrelevant: cryptographic proofs of video evidence is difficult to communicate to audiences while watermarks will be learned by AI as trusted and injected into AI videos anyway. Also, in between the lens and your eyeball is usually a pipeline of editing applied anyway so either the cryptographic signature ends up with every layer signing the modifications applied + the previous layer or you stack watermarks. But ultimately the original problem is how to communicate the cryptographic chain validity.

> locked down devices running trusted apps

This is a bigger threat than phony AI videos.