Google just made a decision that most AI image tools have been too cautious to make: giving users the choice to turn off visible watermarks on AI-generated content. According to TechCrunch, the change covers images, videos, and songs produced through Google’s AI tools, and it’s rolling out across Gemini and the video editor Flow, with Search support coming shortly after.
The key detail here is what Google is not removing. The invisible SynthID watermark stays. The C2PA metadata stays. So if someone runs a generated image through Gemini or a compatible tool, the provenance data is still there. What disappears is the visible stamp that makes AI-generated content look like a draft or a disclaimer rather than a finished product.
Josh Woodward, Google’s VP for Gemini, framed it as a balance between creative control and safety. That framing is accurate, but it also reflects a real tension the industry has been avoiding. Visible watermarks were always a compromise. They made AI outputs immediately identifiable to the naked eye, but they also made those outputs much harder to use in professional contexts, marketing, editorial, product design, and anywhere polish matters. Removing the visual mark while keeping the embedded metadata is a more sophisticated approach, and honestly, the right one.
The update will apply to Imagen, Veo, and Lyria models. Once available, users can find the toggle under Settings, then Media Watermark. Google is also open sourcing a library called Credentio, which lets developers embed local validation directly into their own apps, a move that extends the transparency infrastructure beyond Google’s own products.
This lands at an interesting moment. Anthropic recently drew criticism for adding watermarks to Claude-generated text and files to meet EU regulatory requirements, a move that sparked debate about whether watermarking text is even technically meaningful or just a compliance gesture. Google’s approach is different in both method and philosophy, and it puts pressure on other providers to think more carefully about what watermarking is actually supposed to accomplish.
For developers building on top of these models, this matters. The visible watermark was often the first complaint from clients or end users. Removing it while maintaining SynthID compatibility means the transparency layer still exists for auditing purposes, but the output can be used as-is. That’s a meaningful improvement. And for the broader conversation around AI content provenance, Google is essentially arguing that invisible standards are more durable than visible ones. That’s probably right.
