Google makes AI watermarks optional. SynthID still marks it.

If you have ever generated a product shot in Gemini and then cropped out the little sparkle in the corner, Google just saved you the step. The visible AI watermark on Google's generated images, video and music is now optional — a setting, not a fact of life. The part worth reading twice: the invisible marking did not go anywhere. You can turn off the badge your customers see. You cannot turn off the one your platform, your marketplace, or a regulator can read.

What actually happened

Josh Woodward, Google's VP for Gemini, confirmed on August 14 that a new Media watermark setting lets users drop the visible mark from output produced by Nano Banana (images), Omni (video) and Lyria (music), inside Gemini and the Flow video editor, with Search support following. He framed it plainly as a tradeoff between creative control and safety: the visible mark is optional now, and invisible SynthID watermarks plus C2PA metadata are still applied (TechCrunch). It rolls out over the coming days.

The day before, Google open-sourced the other half of this. Credentio is a C++ library for validating C2PA Content Credentials, released against spec versions 2.2 and 2.4. Google says the same code already backs close to 40 of its own C2PA-conformant products across tens of billions of generated assets. The design point that matters: it validates entirely locally, so an app can check a file's provenance without shipping the media to somebody's cloud (Google Developers Blog). Today it reads credentials; writing them is on the roadmap.

Why AI watermarks matter for your business

Read those two announcements together and the direction is obvious. The visible sparkle was never the enforcement layer — it was the courtesy. The enforcement layer is metadata riding inside the file, and Google just handed every developer a free, local, fast way to check it.

So plan for a near future where the platforms you sell on can cheaply sort AI-generated media from camera-captured media at upload time. Amazon already asks sellers to tag AI-generated people in product images. The EU AI Act's Article 50 transparency rules landed August 2. Turning off a visible badge does not opt you out of any of that; it just means you can no longer see, at a glance, which assets in your own library are generated.

That is the practical risk, and it is boring: your asset pipeline stops self-documenting. Fix it on your side. Record generation provenance in your DAM or product database when the asset is created — model, prompt, date, operator — instead of relying on a pixel in the corner to remind you six months later. And do not let "the watermark is gone" become internal shorthand for "this is indistinguishable from a photo." It isn't. It's just quieter.

Sources: TechCrunch, Google Developers Blog.