Google announced on August 14, 2026 that users can turn off the visible watermark on AI images, videos, and songs made with its models. The invisible SynthID mark and C2PA metadata stay in place.
The Big Picture
- Google will let users switch off the visible spark icon on images, videos, and songs from its Gemini models.
- Josh Woodward, Google’s vice president for Gemini, said invisible SynthID marks and C2PA metadata remain on every file.
- SynthID has watermarked over 10 billion pieces of content, and the change does not affect that layer.
- The toggle sits under Settings, then Media watermark, and it stays switched on by default.
- Google is also open-sourcing Credentio, a library that lets developers check content credentials inside their own apps.
Google makes the corner spark icon optional
Google is rolling the setting out across Android, iOS, web, and macOS over the coming days. Users will find a new “Media watermark” section in Settings with plain On and Off options, and On remains the default. The toggle covers images from Nano Banana, video from Omni, and songs from Lyria.
Google is shipping the same control in Flow, its AI video editor, and says Search will get it next. The company is withholding the option in countries where law requires the visible mark to stay. It has not published that list.
The control arrived in the same window as Gemini 3.7 Flash, which replaced 3.6 Flash in the model picker and now serves chats as well as the Spark agent. That model sits alongside 3.5 Flash-Lite and 3.1 Pro, the same picker where the Gemini app usage limits apply.
Rolling out over the next few days, you can decide if your image, video, and music creations made with Gemini will have visible watermarks. Invisible SynthID watermarks and C2PA metadata will always stay embedded in the background.
— Google Gemini (@GeminiApp) August 14, 2026
Simply go to your settings and toggle “Show… https://t.co/b0tHkwnFBU
SynthID and C2PA stay, but the check moves to a tool
Josh Woodward framed the decision as a trade between two goals. He said:
That last sentence carries the real shift. A viewer could spot a Gemini image by the spark icon sitting in the corner of the frame. Once the toggle ships, the same viewer has to push the file through Gemini or Search to get an answer, and Search is still waiting on its own AI Mode rollout.
Google is releasing Credentio, an open-source C2PA library, so developers can run that validation locally inside their own apps.
Anthropic moved the other direction this week
Anthropic added watermarks to text and files generated by Claude days earlier, citing EU regulatory compliance. Some Claude users objected in public, saying the marks would expose work they had presented as their own at jobs and in classes.
The two calls bracket the same problem from opposite ends. Google is loosening the label a person can see on synthetic media, while Anthropic is stamping one onto text that previously carried nothing at all.
What the announcement does not answer?
Google’s statement establishes that the invisible layer stays embedded at the moment of generation. It does not establish that the layer survives what happens to a file afterward, and Google published no durability data with the change. Four questions sit open:
- Which countries keep the visible watermark mandatory under local law?
- Does SynthID survive cropping, screenshotting, or re-encoding?
- When does the toggle reach Search?
- Will third-party platforms read C2PA credentials on upload?
Why It Matters?
The change answers a real complaint from people who publish AI media for a living, where a fixed corner icon can make an otherwise usable asset unusable in a layout. It also moves the work of identifying synthetic media onto the person looking at it. Verification now runs through Google’s own products, which concentrates the answer inside the company that generated the file, and the stakes show up in misinformation research and in AI-driven fraud threats.
Anyone who has been treating the spark icon as a reliability signal should stop once the rollout completes, because its absence will no longer mean a human made the image. Running a suspect file through Gemini or Search helps reduce the risk of passing synthetic media along as real, though no detection layer catches everything. Teams that republish user-submitted images and video can preserve C2PA metadata during processing instead of stripping it, which keeps the provenance trail readable for whoever checks next.
