Google Will Let Users Remove Visible Watermarks From AI Generations

Google announced it will allow users to remove the visible watermark from its AI-generated images, videos, and songs, rolling out a toggle in the coming days across Gemini and the Flow video editor, with Search support following soon after. TechCrunch

Josh Woodward, Google's VP for Gemini, said the visible-watermark toggle will be available for the Nano Banana, Omni, and Lyria models. Users will be able to turn the visible watermark on or off via Settings > Media Watermark. The setting will be accessible in Gemini and Google's video editor Flow, with support for Search coming soon.

Crucially, removing the visible watermark does not affect the invisible SynthID watermark or C2PA standard-related metadata that Google embeds in its AI-generated outputs. SynthID adds an invisible digital watermark to AI-generated images, videos, and audio without degrading quality, and Google's SynthID Detector can help determine whether an image created using Google AI contains that watermark. The Gemini app also uses SynthID to verify whether an image was made or edited by Google AI.

Google is additionally open-sourcing a library called Credentio, which lets developers embed a local C2PA content-credentials validation mechanism in their applications. Credentio is designed to make content provenance verification accessible outside Google's own products.

The broader context for this move involves multiple threads. At Google I/O 2026, Google announced an expansion of SynthID and OpenAI disclosed that it is adopting the watermarking standard. The original commitment traces back to July 2023, when Google, OpenAI, and other companies pledged to watermark AI content for safety in a White House announcement. Google is also maintaining a visible watermark — the Gemini sparkle — on images generated by free-tier and Google AI Pro-tier users of Nano Banana Pro, in addition to SynthID, suggesting the new toggle does not apply uniformly across all product tiers.

More recently, Anthropic announced it would include a watermark in text and files generated by Claude to comply with EU regulations. Google's decision to make its visible watermark optional moves in the opposite direction from Anthropic's compliance-driven approach, while leaning harder on the invisible SynthID layer as the primary provenance mechanism.

The tension here is between user control and content authentication. Visible watermarks serve an immediate signaling function: a person encountering an image in a feed or on a platform can tell at a glance that it was AI-generated. SynthID, by contrast, requires tooling to detect. Ars Technica reported in late July that simply cropping SynthID-watermarked images can often eliminate the watermark, raising questions about how robust the invisible layer is in practice. If users remove the visible mark and the invisible mark is vulnerable to routine manipulation, the combined provenance system has a meaningful gap.

Worth flagging is the architectural bet Google is making. By open-sourcing Credentio and expanding SynthID's adoption — including OpenAI's integration — Google is positioning invisible, metadata-level provenance as the industry standard rather than visible labels. The C2PA content-credentials standard provides a structured way to assert provenance, and a local validation library lowers the barrier for third-party apps to participate. The question is whether invisible watermarks, however well-designed, can survive the adversarial environment of real-world image distribution, where crops, re-encodes, and platform-level stripping are routine.

Google's announcement also lands alongside a busy product cycle. The company announced Gemini 3.7 Flash, which it describes as its most intelligent workhorse model yet for coding and agents, and published its Made by Google 2026 collection of Pixel device announcements. The watermark toggle, Credentio, and model updates together suggest Google is consolidating its AI provenance and generation tooling into a more coherent platform proposition.

The pragmatic read is that Google is acknowledging what has been true since visible watermarks were introduced: they are a friction point for creators who want to use AI-generated content in polished contexts, and they are trivially separable from the underlying asset in any case. Doubling down on SynthID and C2PA metadata, while giving users control over the visible mark, is a reasonable hedge — provided the invisible layer holds up under the kind of casual manipulation that Ars Technica's testing exposed. That provision is doing a lot of work, and the next round of independent SynthID robustness testing will tell us whether Google's two-layer approach is structurally sound or just more convenient.