One Toggle Removes the Gemini AI Watermark, Not the Proof

Google has made the Gemini AI watermark optional, so you can now strip that small sparkle off pictures, video and music the app generates for you. The switch lives in the Gemini app's settings, under Media Watermark, and it works inside Flow, Google's video editing tool, as well. Support in Google Search is meant to follow. What the toggle does not do is make the file anonymous.

Josh Woodward, Google's vice president for Gemini, announced the change on 14 August and framed it as a trade-off rather than a retreat. Output from three model families is covered — Nano Banana, Omni and Lyria, which between them handle images, video and audio. Those are the ones that have been stamping every generation in the corner. The sparkle goes. Two invisible layers stay, and they travel with the file wherever you send it.

That distinction matters more here than in most markets. The Ministry of Electronics and Information Technology notified amendments to the IT Rules on 10 February 2026, and the synthetic media provisions took effect on 20 February after a ten-day compliance window. Under them, AI-made content is supposed to carry a clear disclosure plus durable provenance data. So an Indian creator isn't only picking a look — they're deciding what a platform sees at upload.

What Google Actually Turned Off

Only the cosmetic bit. The visible Gemini AI watermark was always the flimsiest part of the labelling stack — one screenshot with the corner cropped and it was gone, which is precisely why Google never leaned on it alone. Switching it off is closer to hiding a TV channel's logo bug than to filing off a serial number. Everything the detection tools actually read is untouched.

The Gemini AI Watermark Is Only the Top Layer

Beneath the sparkle sit two separate systems doing different jobs. One hides inside the pixels and the audio itself. The other rides along in the file's metadata. Between them they answer two questions a platform or an editor will eventually ask: was a machine involved, and what has happened to this file since it was made?

SynthID, the mark that survives a screenshot

SynthID buries a signal in the content at a strength no human eye or ear can register. Google says it holds through screenshots, crops and format conversion — the everyday rough handling that wipes out a visible badge. Media can be run through the SynthID Detector to check whether a Google model produced it, and that check works the same whether you kept the sparkle or not.

C2PA credentials, the file's paper trail

The second layer is Content Credentials, the C2PA standard that Adobe, OpenAI, several camera makers and a number of news agencies have signed up to. It records how a file was created and edited in a signed block attached to the file. Google has added these credentials to Nano Banana Pro, so one image can now carry both a hidden signal and a tamper-evident history.

Credentio, the checker Google gave away

A day before the watermark news, Google open-sourced Credentio, a C++ library that validates those credentials entirely on the device, with nothing uploaded to a server. It's the same code behind close to 40 C2PA-enabled Google products, which the company says have processed tens of billions of files. For a newsroom or fact-checking desk in India, that means provenance checks can run in-house, on a laptop, without shipping sensitive material anywhere.

Where India's Labelling Rules Come In

The February amendments were written with deepfakes in mind, not people making Diwali posters. An earlier October draft had proposed a hard size rule: a label covering at least a tenth of the visual area, or the first tenth of an audio clip. That numeric threshold was dropped from the notified version, according to MediaNama, replaced by a looser requirement for a clear, continuous disclosure. The obligation sits mainly with platforms, which is why hiding the sparkle doesn't put you personally in breach — it just hands the labelling job to Instagram, YouTube or X, which will read the metadata you kept.

Why Creators Wanted the Sparkle Gone

Talk to anyone running a small design studio or selling on Instagram and the grievance is identical: a client-ready thumbnail with a vendor's mark burned into it looks like a draft. Competition played a part too. Plenty of rival image tools ship clean files on paid plans, and a permanent badge became a reason to leave. There's a consistency argument as well — a heavily retouched camera photo carries no badge, while an AI illustration of a chai stall did. As Woodward put it:

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.

What to Check Before You Post

Treat the new setting as a design choice, not a privacy feature. The Gemini AI watermark toggle is rolling out over several days, so it may not appear for everyone at once. A few things worth doing once it does:

• Open Settings in the Gemini app and look for Media Watermark. Flow has the same control; Search is still pending.
• Assume the file still identifies itself. If you'd be uncomfortable with a client, employer or platform knowing it was generated, the toggle will not cover you.
• Tick the AI-content declaration where a platform offers one. Most now do, and misdeclaring is what actually costs people their reach or their account.
• Don't treat re-exporting as laundering. Heavy recompression can strip metadata, but SynthID is built to survive exactly that.

Watch what Indian platforms do next. If they begin auto-labelling uploads by reading embedded credentials rather than waiting for a user to tick a box, the sparkle becoming optional will barely register — the label will simply reappear once your post goes live, this time written by someone else.