Google lets Gemini and Flow users switch off visible AI watermarks
Google is rolling out a setting that lets Gemini and Flow users disable a visible watermark on generated media while retaining invisible SynthID watermarks and C2PA metadata. The change makes the most immediate disclosure cue optional, with an exception for Flow users in India, South Korea and Vietnam.
- Google is rolling out a Gemini and Flow setting that can disable a visible AI watermark.
- Google says invisible SynthID watermarks and C2PA metadata will remain, and it is open-sourcing a validation library.
- The switch is not universal: Flow automatically applies the visible mark for users in India, South Korea and Vietnam.
Google is making its visible label for AI-generated media optional. The company says a new setting in Gemini and Flow will let users turn the visible “sparkle” watermark on or off, while leaving invisible SynthID watermarks and C2PA metadata in place.
Josh Woodward, the Google vice president responsible for Google Labs, Gemini and AI Studio, announced the policy change. His post says it covers Nano Banana images, Omni videos and Lyria songs, and that Search will be next. The setting is rolling out in the coming days; users will find it at Settings > Media Watermark once it reaches their account, a report on the rollout says.

Google’s company-reported Gemini menu shows the Media watermark set to Off. Source: Josh Woodward on X.
Google is separating visible disclosure from provenance
The toggle does not mean Google is abandoning identification of generated material. Woodward said the company will continue to use SynthID and C2PA metadata “for transparency” and that Gemini or Search can be used to check whether an image was AI-generated. In other words, the change makes the cue that a person can see at a glance optional while retaining machine-readable provenance signals.
That distinction is the point of the release. The visual mark can interfere with creative or professional use, as the reporting notes, but it also offers a quick disclosure that an invisible marker does not. A later check can establish that a file carries a marker; it cannot give a casual viewer the same immediate signal.
The visible mark was never a complete safeguard. The report says users were already editing visible watermarks out. It also notes that OpenAI and Meta image generators do not apply visible watermarks, while Anthropic had announced invisible watermarks for AI-generated text and images. Those systems are not interchangeable, but the comparison narrows Google’s change: it removes a visible practice that was not the common industry baseline.
