Google lets you switch off the visible AI watermark on Gemini songs, and SynthID stays regardless

Google made the visible AI watermark optional on everything Gemini generates, songs included. The Media Watermark setting landed on August 14, 2026, and TechCrunch reported it covers the Nano Banana, Omni and Lyria models. Lyria is the one that writes music, so a track made in the Gemini app can now leave without a mark anyone can see.

The invisible layer does not move. Google says SynthID and C2PA metadata stay embedded either way, and SynthID survives screenshots, crops and format conversions.

What Gemini’s Media Watermark toggle changes for AI songs

The setting sits in Gemini’s Settings menu in the left bar, near the bottom, with two states: On and Off. Android Authority found it as the third-to-last item in the menu. It also works in Flow, Google’s video editor, and Google says Search support is coming.

Josh Woodward, the Google VP running Gemini, AI Studio and Google Labs, had asked users for their biggest complaints about the product. Stripping the visible watermark came back in the top 10. So this shipped as a feature request, not as a policy shift.

“

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.

Google is also open-sourcing a library called Credentio for working with content credentials, which lines up with the C2PA half of the claim.

Why platforms labelling AI music now depend on a mark you cannot see

Follow the timing. EU AI Act transparency rules took effect this month and put the marking duty on the companies building the models. Suno added audio watermarking to its own output in early August. Anthropic started marking Claude text on August 13. Google made its human-readable mark optional the next day.

Nothing here is a violation. The EU duty leans on machine-readable marking, and SynthID is exactly that. What changes is who can perform the check. A DJ, an A&R, or a playlist curator could read a visible watermark in a second. Reading SynthID takes a query to Gemini or Google Search, which means the verification step now belongs to whoever bothers to run it.

That lands awkwardly on the platforms writing AI policy right now. Spotify’s AI Persona badge depends on artists self-declaring, and Beatport withholds AI uploads at ingestion using Beatdapp detection. Neither reads a visible sparkle. Both are built for a world where the label travels invisibly or not at all.

For your own releases the practical answer has not changed. Your distributor’s upload form still has a disclosure field, and that field, not Google’s toggle, is what platforms read when they decide how to treat your track.