Google makes Gemini’s visible AI watermark optional in three models

Google is letting you switch off the visible watermark on images, video and music generated in Gemini. The company announced the change on Friday, and TechCrunch reported it covers the Nano Banana, Omni and Lyria models. The invisible marks are staying.

The control sits under Settings, then Media Watermark, in the Gemini app and in Flow, Google’s AI video tool, with Search support coming. Turning it off strips the sparkle that has been sitting in the bottom right corner of generated files. Every file still carries a SynthID signal and C2PA provenance metadata, so you can ask Gemini or Search whether a piece of content was machine made.

We’re striking a balance here between creative control and safety: while the visible watermarks are now optional, invisible SynthID watermarks and C2PA metadata are still being used for transparency.

Josh Woodward, VP of Google Labs, Gemini and AI Studio, via TechCrunch

That balance has a geographic limit. The Verge reports the toggle won’t launch in countries whose law requires a visible mark, and notes that plenty of people were already editing the sparkle out by hand. Which made the visible layer a speed bump rather than a control.

The timing is what makes this interesting, because Anthropic spent the same week walking the other way. Claude now embeds invisible marks in text it generates or merely edits, a decision that landed badly with a chunk of its own users, as we covered last week. Google is subtracting a mark you can see and Anthropic is adding one you cannot. Both land in the same place: provenance that only a detector can read.

Provider	Visible mark on media	Machine readable signal
Google	Optional from 14 August 2026	SynthID and C2PA
OpenAI	None	SynthID and C2PA
Meta	None	Content Seal
Anthropic	None	Text watermark, C2PA on files

Anthropic’s half of this is compliance work rather than product design. Providers whose outputs are not detectable face fines of up to 15 million euros, roughly 17 million dollars, or 3 percent of global annual turnover, Nature reported. Models released after 2 August have to meet the requirement immediately, while versions already on the market have until 2 December. Our explainer on the EU AI Act sets out who else the rules reach.

The catch is what invisible provenance costs an ordinary reader. A sparkle in the corner is checkable by anyone with eyes, while SynthID and C2PA need a tool that most people don’t have. Ars Technica notes that a screenshot or a decent metadata editor strips C2PA cleanly, and Nature reports that text marks can disappear under paraphrasing or a rewrite through another model.

The only layer an ordinary reader could verify unaided is the one being switched off.

Rundowns AI

Google is at least shipping something on the verification side. It’s open sourcing a library called Credentio so developers can validate C2PA Content Credentials inside their own apps, locally, without posting media to a server. Anthropic has said it plans a text detection API, though it hasn’t shipped one, and Ars asked for a timeline without getting an answer.

Even so, the marks do catch something. Nature reported that organisers of ICML 2026 watermarked papers sent out for peer review and caught 506 reviewers breaking a no AI policy in one of the conference’s two review streams. That’s the honest scope of the technology: it catches copy and paste, not people working to evade it. The same gap runs through our piece on why AI text detectors keep failing.

What to watch is whether detection catches up with marking. Google’s toggle reaches Search in the coming days, Credentio needs developers to actually adopt it, and Anthropic’s detection API remains a promise. Until those tools are public and independently tested, the answer to whether a file came out of a model sits with the companies that built the model.