Google Lets Users Remove Visible AI Watermarks, Keeps SynthID

Google is making the visible watermark on its AI-generated content optional, a notable shift in how the company handles media provenance. The change lets creators strip the small "AI-generated" mark from images, videos, and songs produced by Gemini models, while keeping the invisible transparency signals that platforms rely on intact.

What happened

Google announced on Friday that users will be able to turn off the visible watermark on AI generations across images, videos, and songs. Josh Woodward, Google's VP for Gemini, said the toggle will be available for the Nano Banana, Omni, and Lyria models, and can be managed from Settings > Media Watermark in the Gemini app and Google's video editor, Flow. Support for Search is coming soon.

Critically, the change does not touch the invisible SynthID watermark or C2PA metadata. Those provenance signals stay embedded regardless of whether the visible mark is shown, so anyone can still check whether a piece of content was AI-generated.

Why it matters

The move is a deliberate balancing act between creative control and transparency. Visible watermarks often make content less useful for professional and creative work, but regulators and platforms still need a reliable way to identify synthetic media. "We're striking a balance here between creative control and safety," Woodward wrote. "While the visible watermarks are now optional, invisible SynthID watermarks and C2PA metadata are still being used for transparency."

The timing is telling. Google's announcement lands right after Anthropic began embedding watermarks in text and files generated by Claude, a move driven by EU transparency rules. Where Anthropic is adding visible markers, Google is making them optional and leaning on invisible, machine-readable signals instead.

Google is also open sourcing a new library called Credentio, which lets developers embed a local validation mechanism in their own apps.

Builder impact

For builders, this splits the provenance stack into two layers: visible marks for humans, now optional, and invisible metadata for machines, still mandatory. If you ship AI-generated media in a product, the takeaway is to lean on C2PA and SynthID-style signals rather than a visible badge, which may soon disappear from the mainstream tools your users rely on.

Read our analysis on Anthropic's Claude watermarks for the other side of this provenance debate.