Google Lets Gemini, Flow Users Turn Off Visible AI Watermarks

Google flipped the default on watermarking this week. Users of the Gemini app and the Flow video editor can now open Settings > Media Watermark and switch off the visible badge that has been baked into images, videos, and songs produced by Nano Banana, Gemini Omni, and Lyria, TechCrunch's Ivan Mehta reported on August 14. Support in Search is described as coming soon.

Josh Woodward, VP for Gemini, made the case that this is not a step back on provenance: 'while the visible watermarks are now optional, invisible SynthID watermarks and C2PA metadata are still being used for transparency.' The pixel-level tag stays, only the human-visible one becomes a user choice. Google also open-sourced Credentio, a C2PA library aimed at letting developers do local validation of that provenance metadata inside their own apps.

The trade is straightforward if you are a creator. Your Gemini output stops looking like Gemini output, which unblocks a lot of client work and social publishing. The trade is less straightforward for everyone downstream. Feeds, moderators, and teachers who had gotten used to that little corner mark as a fast visual tell now have to trust that platforms are actually reading SynthID and C2PA, a bet that only pays off if the invisible signals survive the re-encoding, screenshotting, and cropping that content goes through on the way to a viewer. TechCrunch's write-up does not get into how robust SynthID is to those transformations, and it does not say whether enterprise or business admins can gate the new toggle for managed accounts.

This lands in a busy stretch for Google's AI stack. The Gemini app recently crossed a billion users, and Credentio is the second open-source drop from the company in as many days, arriving alongside the HEIR compiler for encrypted inference. If invisible provenance actually holds up in the wild, this becomes the template for how the rest of the industry stops branding creative output. If it does not, this is the week AI content stopped being labeled at the point where people actually see it.