Google Keeps Invisible SynthID and C2PA Watermarks in All Gemini AI Images, Video and Music

Picture this: you use Google’s Gemini to generate an image for your business’s social media page, then switch off the visible “AI” corner label so it looks cleaner. The watermark you can see is gone. But the one you can’t see? That’s staying put — permanently baked into the file.

Google has confirmed that its invisible SynthID watermarks and C2PA provenance metadata will remain embedded in all AI-generated images, videos and music produced through Gemini, regardless of whether users choose to hide the visible watermark. The Verge reported the detail after Google introduced a new setting allowing users to toggle off the visible AI label in Gemini outputs — a change that, on the surface, sounds like more creative freedom, but comes with a clear caveat.

What SynthID Actually Does

SynthID is a digital watermarking technology developed inside Google DeepMind. It works by embedding imperceptible signals directly into AI-generated content at the moment of creation — signals that Google’s systems can later detect to determine whether a piece of media was made or edited using its AI models.

It’s not a logo or a label. You won’t see it, hear it, or notice it in any conventional sense.

The technology now covers four media types: images, audio, video and text. On the image side, Gemini’s Nano Banana model carries SynthID watermarks by default. Video outputs from Omni and music generated by Lyria do too. SynthID Text, meanwhile, is available to developers building on Google’s platforms.

And the watermark is designed to survive. Google says it holds up through common edits — cropping, filtering, frame rate changes, lossy compression. A Google spokesperson defended the system publicly, describing SynthID as a watermarking solution designed to be difficult to remove without considerably degrading media quality.

The C2PA Layer

Alongside SynthID, Google also embeds C2PA metadata in its Gemini outputs. C2PA — the Coalition for Content Provenance and Authenticity — is an industry-wide framework that attaches cryptographically signed records to digital files, describing how and by whom content was created or edited.

Think of it as a digital paper trail. Where SynthID is Google’s own proprietary signal, C2PA is an open standard developed by a coalition of technology companies and media organisations. Together, they give AI-generated content two separate layers of machine-readable provenance.

Both remain in every file, whether the visible watermark is on or off.

Why Google Is Keeping Them

Google frames the decision around transparency and trust. The company has linked SynthID and C2PA to its commitments under emerging AI regulatory frameworks, including transparency codes of practice connected to the EU AI Act.

It’s worth being clear about one thing: in countries where visible AI labelling is legally required, users won’t be able to turn off the visible watermark at all. The new optional toggle only applies where local law permits it.

The broader picture here is an industry moving towards content provenance as a response to deepfakes, misinformation and the growing difficulty of knowing what’s real online. Google isn’t alone — major technology companies across the sector have been moving towards C2PA adoption, and regulators in both the UK and EU have pushed for stronger transparency requirements around AI-generated media.

But not everyone is entirely comfortable with the approach.

Privacy and civil liberties advocates have raised questions about what pervasive watermarking and provenance metadata could mean for tracking or profiling creators if those systems aren’t strictly governed. Some technical researchers also question whether watermarking alone provides genuine security, pointing out that determined actors may find ways to corrupt or remove marks — potentially creating a false sense of confidence in content verification. And some creators argue that proprietary watermarking systems give technology companies too much control over how AI-generated work is labelled, calling instead for open, user-governed standards.

How Verification Could Work for Everyday Users

Google has already begun integrating SynthID checks into consumer products. Through Google Search and Chrome, users can inspect whether an image or video carries AI-generated markers — a feature that could become increasingly useful as AI-generated media spreads across social platforms and news feeds.

Journalists, educators and public sector communicators are among those likely to benefit most from reliable provenance signals. The ability to run a quick check on a suspicious image or video, and receive a machine-readable answer about its origins, is a practical tool in the fight against misinformation.

For professional artists, musicians and videographers, the picture is more mixed. Some welcome the ability to distinguish human-made work from AI outputs. Others worry about stigma attaching to AI-assisted creativity, or about the implications of having their creative process recorded in metadata they don’t fully control.

What This Means for Kent Residents

Anyone in Kent using Google Gemini to create images, videos or music — whether for a small business, a school project, or a community group’s social media — will have invisible SynthID watermarks and C2PA metadata embedded in their files by default, even if they choose to hide the visible AI label. Local creators and organisations should be aware that this machine-detectable provenance information travels with the content wherever it’s shared. Schools, colleges and public sector bodies such as Kent County Council may find these built-in credentials a useful tool for identifying AI-generated material in safeguarding, digital literacy and communications work. And if future UK legislation mandates visible AI labelling in certain contexts — something both Westminster and Brussels are actively considering — Kent users may find the option to hide visible watermarks restricted or removed entirely.