Google separates presentation from provenance

Google says users will soon be able to remove the visible watermark from AI-generated media created with its tools, including images, videos, and songs. The company is not removing all provenance signals, however: invisible SynthID watermarks and metadata tied to the C2PA standard will remain in place.

The change reflects a practical tension in generative AI. A visible watermark is easy for people to notice, but it can also make a piece of generated media harder to use in professional or creative workflows. Google’s new approach gives users more control over the finished look of their outputs while keeping machine-readable signals that can help identify AI-generated files.

Where the setting will appear

Josh Woodward, Google’s vice president for Gemini, said in a post on X that the toggle will be available for the Nano Banana, Omni, and Lyria models. The setting will appear in Gemini and in Flow, Google’s video editor, with Search support coming soon.

The key facts are straightforward:

• Media covered: AI-generated images, videos, and songs;
• Models named: Nano Banana, Omni, and Lyria;
• Initial products: Gemini and Flow;
• Search: support is coming soon;
• Control path: Settings > Media Watermark;
• Rollout: expected over the coming days.

Once the feature is available, users will be able to switch the visible media watermark on or off from that settings menu. For creators, the immediate benefit is cleaner output for editing, publishing, or presentation. For Google, the challenge is to preserve trust signals even when the most obvious label is no longer shown.

Hidden signals stay in the file and ecosystem

Google emphasized that turning off the visible watermark will not affect SynthID or C2PA-related metadata. SynthID is Google’s invisible watermarking system, designed to help identify AI-generated media without placing a visible mark on the surface of the content. C2PA is a provenance metadata standard that can carry information about where a file came from or how it was created.

Woodward framed the decision as a balance between creative control and safety. In his words, visible watermarks are now optional, but invisible SynthID watermarks and C2PA metadata will still be used for transparency. He also said users can still use Gemini or Search to see whether an image was AI-generated.

That distinction matters. Visible labels are useful for immediate disclosure, but they are also intrusive and may be cropped, obscured, or simply unsuitable for final assets. Invisible watermarks and metadata are less obvious to viewers, but they are better aligned with automated verification, platform checks, and search-based provenance tools.

Google is also open sourcing a new library called Credentio. The stated goal is to let developers embed a local validation mechanism in their own apps. In practical terms, that gives software makers another way to check provenance signals without relying only on what a user can see on the media itself.

A broader shift in AI labeling

The announcement arrives as major AI companies continue to rethink how generated content should be labeled. Anthropic recently drew debate for adding a watermark to text and files generated by Claude in order to comply with EU regulations. Google’s move points in a different direction: make the visible mark optional, but keep background verification intact.

For the AI industry, this is likely to be an important middle path. Professional users want AI-generated assets that can fit naturally into creative work. Platforms, regulators, and audiences still need ways to understand whether content was generated or altered by AI. The long-term answer may not be a single visible badge, but a stack of signals: hidden watermarks, provenance metadata, search tools, and developer validation libraries.

Google’s update suggests that AI disclosure is becoming less about stamping every output with an obvious mark and more about building verification into the surrounding ecosystem. If that model holds, the next competitive frontier will be whether these provenance systems are reliable, interoperable, and easy for ordinary users and developers to verify.