Provenance by default

Optional watermarking is theatre. Google shipping SynthID on by default across YouTube-scale distribution is the only version of this that could work.

Gemini Omni applies SynthID by default, across a model unifying text, image and video generation that ships to the Gemini app, Google Flow, YouTube Shorts and YouTube Create.

The default is the entire policy

An opt-in watermark marks the content of people who were never going to mislead anyone. The people you care about switch it off. Default-on inverts that: everything is marked except what somebody deliberately stripped, and stripping acquires intent.

Combine it with distribution at YouTube scale and a meaningful fraction of consumer-generated video carries provenance without a single user making a decision. Defaults are the only mechanism that has ever delivered a safety property at population scale.

What it does not do

Watermarking is not tamper-proof. Re-encoding, cropping and adversarial processing degrade it; a determined actor removes it. And it says nothing about models that do not implement it — which is most of them, including the open-weight video systems now within a few Elo of the top of the board.

This is not a solution to synthetic media. It is infrastructure for the ordinary case, which is most cases.

The compliance angle nobody is stating

Colorado's chatbot law puts disclosure duties on the operator, and two dozen states moved similar bills. Disclosure obligations need an evidentiary basis. A platform that can already demonstrate which of its outputs were machine-generated is holding the artefact those rules will ask for.

Which suggests default-on was not only a safety decision. The best compliance posture is usually the one you adopted before anybody required it.

The Next Web — Google launches Gemini Omni Flash, a conversational video-generation model → · Digital Applied — Gemini Omni: Multimodal Video Creation Guide for 2026 → · Google AI for Developers — Gemini API release notes →