Google Makes Its AI Watermark Optional, but Invisible Ones Stay

Key points

• Google added a toggle in August 2026 that lets users remove the visible sparkle watermark from AI-generated images, videos, and music.
• Invisible SynthID watermarks and C2PA metadata remain embedded in all AI-generated content, even with the visible mark turned off.
• Google's Nano Banana and Omni AI models are affected by the change.
• OpenAI and Meta already skip visible watermarks on their AI image tools, relying on hidden standards instead.
• Anthropic announced the same week that it's adding invisible watermarks to both AI-generated text and images.

Google quietly handed users a new option this week: you can now turn off the small sparkle icon that appears in the corner of images and videos made with its AI tools. The toggle lives in the settings of Gemini and Flow, its AI video generator.

Flip it off and the visible badge disappears. What stays put is something most people will never see.

What are these invisible watermarks, exactly?

Every piece of AI-generated content still carries two hidden markers. First, SynthID, a signal Google's DeepMind lab bakes directly into the pixels or audio of the file itself, which survives most cropping and resizing. Second, C2PA metadata, a standardised record attached to the file that logs where and how the content was created, similar to the data a camera embeds in every photo it takes.

Neither marker is visible. But tools built to read them, Gemini and Google Search among them, can detect both. Ask either whether an image is AI-generated and it checks for those signals.

Josh Woodward, Google's vice president of Labs, confirmed both layers remain active regardless of the visible watermark setting.

Does losing the visible mark make deepfakes easier to spread?

Possibly, though it's more complicated than it first looks.

Many users have already been editing the sparkle out manually, so the visible mark was never a reliable safety net. OpenAI's image tools never applied visible watermarks, relying on SynthID and C2PA instead. Meta launched its own standard, called Content Seal. Anthropic announced this week that invisible watermarks now cover both the text and images its models produce. The industry has effectively settled on a hidden-watermark model, which puts the burden of detection on software rather than on a logo anyone can crop out in seconds.

For ordinary users the practical change is small but real. A visible sparkle let a casual viewer spot AI content at a glance. Without it, that instant cue is gone. The technical trail is still there, but reading it requires a tool. We covered the C2PA standard as far back as 20 July 2026, and our reporting on Apple's verification feature shows how that same infrastructure is being built into device cameras too.

Google plans to roll the toggle out to Search as well. It won't appear in countries whose laws require a visible mark on AI-generated content.

What should you watch for?

Visible watermarks are gone from most major AI image platforms now. You can't rely on a logo to tell you what is real. If you receive an image or piece of audio and you're unsure whether AI made it, upload it to Google Search or Gemini and ask. The hidden signals should still answer the question.

Be especially cautious with images shared on social media after re-saving or heavy compression. Compression can strip C2PA metadata even if SynthID survives. When the stakes are high, treat unverified media with scepticism regardless of how convincing it looks. Neither system is foolproof, and the logo that once made this easy is no longer there.