Google Lets Users Turn Off Visible AI Watermarks on Nano Banana, Flow and Lyria Content

Google confirmed on Friday that people generating images, video and music through its AI tools can now switch off the small visible watermark that has stamped every output since launch. The invisible SynthID layer stays in place regardless of the setting, and so does C2PA metadata attached to files.

Josh Woodward, Google's VP for Gemini, posted the announcement on X rather than through a formal press release, framing it as a direct answer to user complaints about the corner logo interfering with professional and creative work.

What actually changes

The toggle removes the small spark icon Google places in the corner of AI-generated media. It does not touch SynthID, the imperceptible pixel-level or audio-level pattern Google DeepMind built to let its own tools detect AI involvement later. It also leaves C2PA metadata untouched, the signed record that travels with a file and documents how it was made.

Woodward put the trade-off in his own words: the company is separating creative control from the transparency layer that regulators and platforms actually rely on. "While the visible watermarks are now optional, invisible SynthID watermarks and C2PA metadata are still being used for transparency," he wrote, adding that Gemini and Search can still be used to check whether a given image was machine-made.

That distinction matters more than it might look at first glance. A visible watermark is a design element. An invisible one is a detection mechanism. Google is giving up the first while insisting the second still works, which is a bet that most people asking for watermark-free exports care about polish rather than concealment.

Where the switch lives and what it covers

Once the update lands, the setting sits under Settings, then Media watermark, in Gemini's account menu. On the web version, that means opening the gear icon in the bottom-left corner and finding the new section there. A single toggle switches the corner branding off.

Coverage spans three model families rather than one. Nano Banana handles the image side, Omni covers video generation, and Lyria produces the AI music that also picks up the option. Google's video editor Flow gets the same toggle, and Woodward said Search support for the setting is coming, without a specific date attached.

The rollout itself is staged. Google says the feature is arriving "in the coming days," which means the setting may not be visible to every account the moment this runs.

A request that had been sitting in the queue for weeks

This didn't come out of nowhere. Earlier in the summer, Woodward asked Gemini's power users to name their biggest frustrations with the product. Thousands responded, and the team distilled the feedback into a top-ten list. Removing the visible watermark from Nano Banana images landed on that list, and Woodward said at the time his team would look into whether it was feasible.

Signs of that work showed up in the wild before Google made anything official. Teardown reports on a Google app update found references to a "Watermark settings" option sitting near existing account controls, suggesting the feature was already in testing weeks before Friday's post. Developers had also flagged a related complaint on Google's own AI forum, where a Gemini API bug report described the forced star watermark distorting video generated through frame-to-video workflows in Flow, since the model was reportedly treating the logo as an object to preserve rather than an overlay to ignore.

None of that guarantees the fix Google shipped addresses every complaint in that thread. But the sequence, from user survey to leaked settings string to public rollout, gives a reasonably clear paper trail for how this feature came to exist.

Credentio: the other transparency move this week

A day earlier, Google published a separate but related piece of infrastructure. Credentio is a new open-source C++ library built to validate C2PA Content Credentials locally, without sending a file to a server to check its provenance. Google says the underlying code already runs across roughly 40 of its own C2PA-conformant products and has processed tens of billions of images, videos, audio files and documents.

The library currently supports C2PA specification versions 2.2 and 2.4, and it can check content against either the official C2PA trust list or a custom one that a developer defines. Local, in-app validation means a company building a verification tool doesn't need to route user media through Google's servers to get an answer, which cuts both latency and the privacy exposure that comes with uploading a file just to check its history.

For now, Credentio only validates existing credentials. Google says it plans to extend the library to cover credential creation and embedding as the C2PA spec itself evolves, meaning the current release is a read tool rather than a full write-and-verify package.

Two AI labs, two opposite bets

The timing puts Google's move next to a very different decision Anthropic made three days earlier. Anthropic confirmed it will begin embedding invisible watermarks into text generated by newer Claude models, a change tied to Article 50 of the EU AI Act, which took effect on August 2 and requires generative AI providers to mark synthetic content in machine-readable form. Companies that skip compliance risk fines up to €15 million or 3 percent of global turnover.

Anthropic's version works by statistically biasing word choice in a way that's undetectable to a reader but identifiable with the right key, and the company says the mark persists through copy-paste and some editing. It applies globally, not only to EU accounts, and it covers the API, Claude, Claude Code, Claude Cowork and Claude Tag, along with Claude access through AWS, Google Cloud and Microsoft Foundry.

The reaction on Reddit and elsewhere has been mixed at best. Some users argue the watermark exposes people using Claude for work or schoolwork they'd rather keep quiet. Others point out that a tool used to draft something is still a tool, and detecting its use isn't the same as claiming authorship over the result.

Google's announcement runs in the opposite direction on the visible layer while keeping the same logic Anthropic is leaning on for its invisible one. Strip the branding people find intrusive, keep the machine-readable signal regulators and platforms want to check against. Whether that combination holds up under real-world testing, especially given that third-party tools already exist to strip Google's visible star watermark by reverse-engineering its blending method, is a separate question from whether Google intended it to.