Google Makes Its Visible AI Watermark Optional

Google is making the visible watermark on its AI-generated media optional. The corner badge that has been stamped onto output from the Nano Banana, Omni and Lyria models can now be switched off from a single toggle under Settings → Media Watermark, covering both the Gemini app and the Flow video editor, with Search support to follow. The change rolls out over the coming days to everyone who already has access to those tools.

What is not becoming optional is the part users cannot see. SynthID, the imperceptible signal Google embeds in the pixels themselves, stays in every generation regardless of the toggle, as does the C2PA provenance metadata attached to the file. Josh Woodward, the Google VP who announced the change, framed it as “striking a balance here between creative control and safety” — the visible mark goes, the machine-readable ones remain.

That distinction is the whole argument, and it cuts both ways. The visible badge was always the weakest link in provenance: it is the one marker a user can crop out in five seconds, and it is also the one that made the output unusable for anyone doing legitimate commercial design work. Removing it as a default arguably costs less than it looks like, because anyone determined to strip it already could. The counterargument is that the visible badge was the only signal an ordinary person scrolling a feed would ever notice, and SynthID only helps if something in the pipeline is actually checking for it.

Google appears to know that, which is why the announcement came with a second piece: an open-source library called Credentio, aimed at letting developers validate content credentials locally rather than round-tripping to a Google service. Shifting verification to something anyone can embed is the more consequential half of the news, since a provenance standard that only one company can read is not really a standard.

The timing puts Google on the opposite tack from its peers. Anthropic began watermarking Claude’s text output worldwide this month under the EU AI Act’s Transparency Code, and OpenAI adopted SynthID and shipped a public verification tool back in May. All three are converging on invisible, machine-checkable provenance; Google is simply the first to conclude that the visible layer on top was doing more harm to its product than good for anyone’s trust. Whether that holds depends entirely on how many platforms bother to read the invisible one.