SevenTnewS

Google AI

Turning off Gemini's watermark doesn't make it untraceable

AI content from Gemini and Flow can now drop its visible sparkle badge. Invisible SynthID and C2PA markers remain, so verification still works. The catch: it only works for people who think to ask.

Emmanuel Fabrice Omgbwa Yasse AI-assisted

2026-08-23 · 4 min read

Google is handing users the switch that undoes its most visible AI tell. A new "Media watermark" setting in Gemini and the company's AI video generator, Flow, removes the sparkle badge from the bottom-right corner of content produced by its Nano Banana and Omni models. What stays behind is quieter: an invisible SynthID watermark and C2PA metadata embedded in the background of the file. It is not the only Gemini story in motion; a leaked test build has raised questions about a pro model that has not shipped.

The distinction matters more than the toggle itself. Visible markers are read by people; SynthID and C2PA are read by software. Josh Woodward, vice president of Google Labs, Gemini, and AI Studio, says content remains verifiable after the change, and anyone can ask Gemini or Search whether a given image, video, or track was machine-made.

Google follows the crowd that never showed its hand

The update brings Google in line with rivals that skipped the badge entirely. Per The Verge, OpenAI and Meta image generators do not apply visible watermarks. OpenAI relies on SynthID and C2PA to flag AI generation, Meta launched its own Content Seal standard and has kept its release pipeline busy with open-weights models like Muse Glimmer, and Anthropic said the same week it would embed invisible watermarks in both AI-generated text and images.

Google was the outlier in attaching a public label to its output, and its own rationale is partly a concession: many users were already cropping the sparkle out by hand. Making the option official formalizes a workaround that was already standard practice. There is a competitive angle too. A corner badge made Google's output easy to single out in feeds where OpenAI and Meta images carry no visible label at all.

The four approaches, as The Verge describes them:

ProviderVisible watermarkInvisible markers
GoogleOptional "sparkle" badge on Gemini and Flow contentSynthID + C2PA
OpenAINoneSynthID + C2PA
MetaNoneContent Seal
AnthropicNot statedInvisible watermarks on text and images

Laid out side by side, the approaches show what Google is not doing: it is not weakening the provenance chain. C2PA metadata can still be inspected, and the SynthID marker was never visible to begin with, so it cannot simply be lopped off in an editor. What changes is where detection happens: at a glance for no one, on request for anyone who asks.

Practically, the change narrows the distance between Google and its rivals more than it widens the gap with forgers. People who wanted to pass AI content off as human already had a one-click crop as their tool; the badge never survived contact with a screenshot. What the toggle actually removes is the friction for everyone else, the momentary pause a visible badge creates before you decide whether to trust what you are looking at.

The fingerprint only helps if you check it

That is the trade the update asks users to accept. On a fast scroll, an unmarked Gemini image gives no immediate sign it is synthetic; discovery depends on a deliberate query. Invisible provenance only works if people take the extra step to check it. SynthID and C2PA answer any question you think to ask. They do nothing for the content you never think to question.

For people who republish or moderate AI media, the practical change is a move, not a removal: where a corner badge once answered the provenance question at a glance, the answer now comes from a query to Gemini or Search, or from inspecting C2PA fields directly. The verification is still possible, just no longer ambient.

A watermark policy that splits by country

The setting is also not universal. Google plans to roll it out to Search as well, but the toggle will not launch in countries that require visible watermarks. The result is a split standard: in some markets AI content keeps its badge, in others it relies on metadata alone, so the same file can effectively carry different proof depending on where it is viewed.

Google is betting the watermark was policing the wrong audience anyway; people determined to strip it were never going to leave it in place. What replaces it is a verification system that works on demand instead of on sight. Detection gets stronger for people who ask and weaker for everyone who does not, and the settings menu now has the switch to prove it.

Get the tech essentials in 3 minutes every morning

One email, every weekday, with what actually matters in AI and tech.