Google Lets Users Turn Off Gemini's Visible AI Watermarks

Google is rolling out a toggle that lets users strip the visible "sparkle" watermark from AI-generated images, video, and music — while keeping invisible tracking marks intact, the company said Friday.

What's changing

The new "Media watermark" setting appears in Gemini and Google's AI video tool Flow, and can be switched off to remove the sparkle mark that currently appears in the bottom-right corner of content made with Google's Nano Banana and Omni models. According to Josh Woodward, VP of Google Labs, Gemini, and AI Studio, the toggle will also cover the Lyria music model, with support for Google Search "coming soon." The feature is rolling out "in the coming days," and once live, the option sits under Settings > Media Watermark.

The invisible tags aren't going anywhere

Turning off the visible mark does not touch Google's provenance system: SynthID invisible watermarking and C2PA metadata will still be embedded in every AI-generated file. That means anyone can still ask Gemini or Search whether a piece of content was AI-made, even after the visible sparkle is gone. Woodward framed it as a tradeoff: "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."

Why now

Visible watermarks make sense for casual sharing but get in the way of professional and creative work — many users have reportedly already been editing them out manually before this change. Making the marker optional acknowledges that reality rather than fighting it, while leaning on the invisible layer to keep provenance checks possible.

A new open-source piece: Credentio

Alongside the toggle, Google is open-sourcing a library called Credentio, aimed at letting developers embed local validation of content provenance directly inside their own apps — without depending on a Google-hosted lookup. For teams building content pipelines that need to verify whether an asset is AI-generated at the point of ingestion, this is the more consequential part of the announcement: it turns provenance-checking from a Google-only feature into something embeddable elsewhere.

The contrast with Anthropic

The move lands just days after Anthropic's own watermarking rollout on Claude-generated text and files drew backlash — some users complained the visible mark could expose AI use at work or in school, under EU compliance requirements. Google's approach draws the opposite line: visible marking is optional, invisible provenance is mandatory. For builders shipping AI content features, that split — user-facing signal is a preference, embedded provenance is not — looks like the direction other platforms will likely follow rather than Anthropic's harder stance.

What to watch

The rollout timeline ("coming days" for the toggle, "coming soon" for Search) is vague enough that teams relying on watermark detection for moderation pipelines should verify behavior once it's live rather than assume the current sparkle-mark behavior persists. Credentio's actual API surface, once published, will matter more than the toggle itself for anyone building on top of Google's content stack.