Google is giving users a toggle to strip the visible “sparkle” watermark from AI-generated images, videos, and music in Gemini and its Flow video tool. The setting, which applies to content from the Nano Banana and Omni models, removes the only immediately obvious signal that a piece of media was machine-made. Josh Woodward, VP of Google Labs, Gemini, and AI Studio, framed it as a shift toward less intrusive detection — but the timing is awkward. Anthropic announced invisible watermarks this week, and the broader industry has been moving toward stronger provenance standards, not weaker visible markers.
Let’s be clear about what’s actually changing. Google isn’t abandoning watermarks entirely. Invisible SynthID markers and C2PA metadata will still be embedded in every piece of generated content, whether the sparkle icon is there or not. You can still ask Gemini or Search whether something is AI-generated, and those tools can read the hidden signals. The practical difference is that a human scrolling past an image on social media no longer gets that instant visual cue. For years, the sparkle was the one watermarked tell that didn’t require forensic tools to spot.
The company is also rolling the toggle out to Google Search eventually, but with a notable carve-out: countries that legally require visible watermarks won’t get the off switch. That’s a quiet acknowledgment that this is less about user preference and more about competing with rivals who never bothered with visible marks in the first place. OpenAI relies on SynthID and C2PA without any sparkle equivalent, and Meta built its own Content Seal standard. Google was the outlier in slapping a glowing icon on everything.
Here’s what I’m watching: the gap between what’s detectable and what’s actually noticed. Most people never run images through verification tools. The sparkle, imperfect as it was, gave the average person a fighting chance at spotting synthetic media at a glance. Removing it makes the internet feel a little more seamless for creators, sure. It also makes the default state of every feed slightly less trustworthy. If invisible watermarks are the only line of defense, then the real question isn’t whether the technology works — it’s whether anyone will bother to check.
