For two years the small Gemini sparkle in the corner of an AI image was doing a quiet job most creators never thought about: it disclosed, automatically and on every file, that the picture was machine-made. On August 14, 2026, Google made that visible mark optional — a single Settings toggle now removes it from every future image, video, and song you generate in Gemini. The invisible layer did not move: SynthID and C2PA metadata still ride inside every file regardless of the setting, so the origin stays detectable behind the scenes even when the label you could see is gone. That split — visible off, invisible on — is the whole story, and it quietly reassigns a responsibility. The watermark that used to announce "this is AI" for you is now something you have to decide about, per platform, at the moment you publish. This guide explains what a visible AI watermark actually is, exactly what Google changed and did not change, why the two watermark layers exist for different reasons, and the practical disclosure workflow that keeps you compliant now that the automatic label is optional.
A visible AI watermark is a mark stamped onto AI-generated media so a human looking at it can tell it was made by a machine — the small Gemini sparkle in the corner of an image is the one most creators have seen. On August 14, 2026, Google made Gemini's visible watermark optional: a single toggle in Settings now strips the corner mark from every future image, video, and song you generate. For a definition-first treatment of the term and how visible marks behave across platforms, see the glossary entry on visible AI watermarks; this guide is about what the change means and what you should do about it.
The critical detail is what did not change. The invisible layer — Google's SynthID watermark plus C2PA provenance metadata — stays embedded in every file regardless of the toggle. So the announcement is precisely a split: the label you could see becomes optional, and the signal you cannot see stays mandatory. That split matters because the visible watermark was quietly doing a disclosure job for you on every asset, and turning it off does not delete the obligation to disclose — it moves that obligation onto you, at the moment you publish. The rest of this guide is about that handoff.
Two different things get called "watermark," and conflating them is the single most common mistake in this whole topic. A visible watermark is a human-facing disclosure: a logo, a sparkle, a "Made with AI" badge, or a corner label rendered into the pixels so a viewer knows at a glance the content is synthetic. An invisible watermark is a machine-facing provenance signal: a pattern embedded in the pixels or the audio waveform that a person cannot perceive but a detector can read back. SynthID is the best-known invisible scheme; the on-image sparkle is the visible one. They serve different audiences and answer different questions — "can a viewer tell?" versus "can a system verify?"
There is a third layer worth naming because Google ships it too: C2PA Content Credentials, a cryptographic provenance record attached to the file's metadata that logs how the file was made and by what tool. C2PA is robust for images, video, and audio and fragile for plain text, which is a different problem covered in the entry on AI text watermarking. For visual and audio media, the practical stack is now: an optional visible badge on top, an invisible SynthID signal in the content, and a C2PA record in the metadata. Google's change touched only the first of those three.
Google announced the change on August 14, 2026, with a rollout described as happening over the following few days across Android, iOS, web, and desktop. The control lives in the Gemini app under Settings, in a new "Media watermark" section: flip it off and future images, videos, and music you generate come out without the visible corner mark. It is not retroactive to a specific past file in the sense of stripping a signal — it governs what future creations display. The setting applies across Gemini's generative models — Nano Banana for images, Omni for video, and Lyria for music — and Google said the option is available in Gemini and its Flow video editor, with support in Google Search coming after. The full news writeup with sourcing is in the news entry on Google letting Gemini users turn off the visible watermark.
Regardless of the toggle, every file Gemini generates still carries an invisible SynthID watermark and C2PA provenance metadata. Google's VP leading the Gemini app framed it as striking a balance between creative control and safety, and made the point explicitly: because the invisible signals stay, you can still use Gemini or Search to check whether an image was AI-generated even after the visible badge is gone. In other words, Google did not weaken detectability — it removed a cosmetic label and kept the provenance machinery. A creator who reads "watermark removed" as "now untraceable" has misunderstood the announcement; the origin of the file is exactly as detectable as it was the day before, to any system that reads SynthID or C2PA.
The option is not universal. Google said it will not be offered where local law mandates a visible AI label, and reporting indicated that in a handful of countries the toggle is either unavailable or limited to a paid tier, while managed work and school accounts do not get the setting at all and keep the visible watermark by default. If you operate in a jurisdiction with a visible-labeling statute or on an enterprise-managed account, assume the corner mark stays. This is the first concrete sign of a pattern worth internalizing: whether your content shows a visible AI label is increasingly a function of where you are and what account you use, not a single global default.
The visible watermark and the invisible one were solving different problems, and understanding that is what makes Google's move coherent rather than a retreat from safety. The visible mark is a trust-and-branding signal aimed at the viewer. Creators had a legitimate complaint about it: a persistent corner sparkle clutters otherwise-usable output, brands the content as Google's rather than theirs, and makes Gemini's images look categorically "AI" in a feed next to competitors whose output carried no such badge. For a professional shipping client work or building a brand feed, that visible mark was a real friction, and it was easy to remove anyway by cropping or a quick edit — so it was doing weak disclosure work at a real usability cost.
The invisible layer is the one that does the durable provenance work. SynthID survives cropping, screenshotting, re-encoding, and mild edits that would erase a corner badge instantly, and C2PA carries a signed record of origin. Because those two persist, the actual safety function — can a platform, a journalist, or a detection tool establish that a file is AI-made — is unaffected by the visible toggle. So the defensible reading is that Google kept the layer that reliably answers "is this AI?" and made optional the layer that was mostly cosmetic and trivially defeated. It is closer to honest than it first sounds: a visible badge you can crop away in five seconds was never a serious provenance mechanism, and pretending it was is worse than removing it and being clear about where the real signal lives.
Here is the practical core. For two years, the visible watermark meant that anyone who generated an image in Gemini published something that was self-disclosing — the badge told the audience "AI" whether or not the creator said a word. Make that badge optional and, for everyone who turns it off, that automatic disclosure disappears. What does not disappear is every platform rule and every law that requires you to disclose synthetic media. The visible watermark was never your legal or platform disclosure; it was the generator's courtesy. Removing it does not remove the obligation — it removes the thing that was accidentally satisfying part of it for you.
Concretely, three separate regimes still apply to your published post no matter what the corner of the file looks like. Platform disclosure: YouTube requires you to flag realistic altered-or-synthetic content, and Meta and TikTok have their own AI-content labels and detection — the compliance detail is in the guide on YouTube's AI disclosure and likeness rules. Law: the EU AI Act's transparency rules require marking synthetic audio, image, and video content, which is a deployer obligation that lands on you as the publisher, not just on the model maker. And likeness: if your generated media depicts a real, identifiable person, likeness detection and consent rules are a separate and stricter layer that a missing watermark does nothing to satisfy. None of these care whether Gemini drew a sparkle in the corner. All of them care what you did at publish time.
Stop treating "the file has a watermark" and "the post is disclosed" as the same decision. They are two independent choices. Whether you keep Gemini's visible badge is an aesthetic and jurisdictional call — off for a clean brand feed where the law allows it, on where you want the file to self-disclose or where you're required to. Whether the post carries an AI label is a compliance call governed by the destination platform and your jurisdiction, and you make it in the platform's own disclosure control at upload. A clean, badge-free export can and should still go out with the platform's "AI-generated" label toggled where that's required. Conflating the two is how creators end up either cluttering every asset needlessly or shipping undisclosed synthetic media into a platform that expected a label.
Do not go out of your way to strip SynthID or C2PA. Beyond the fact that SynthID is designed to be hard to remove, scrubbing provenance to pass AI work off as human is exactly the deceptive move platforms in 2026 are built to punish, and it converts a compliant "AI-assisted, disclosed" position into a "tried to hide it" one the moment a detector catches it. The durable stance — the one that survives detection becoming universal — is to keep the invisible provenance, disclose where required, and compete on whether the content is actually good. The broader argument for provenance-you-have-nothing-to-hide-from is in the guide on AI content provenance.
The reliable way to not miss a required label is to make disclosure part of the act of publishing rather than something you remember later. When one asset fans out to many platforms — each with its own rules, its own label control, and its own jurisdiction — the disclosure decision has to be made per destination, and it has to be made by a human who can see the platform's requirement. That is a workflow problem more than a policy problem, and it is where a publishing layer earns its place.
Kompozy is an AI content generation and multi-platform publishing engine, and the watermark change lands squarely on the part of the workflow it owns: the moment a finished asset becomes many published posts. Once the visible badge is optional, disclosure stops being automatic and becomes a per-platform decision at publish time — and Kompozy publishes each destination individually rather than blasting one identical post everywhere. That means the caption, and any AI-disclosure label, is set per platform, so a clean watermark-free image can still ship correctly labeled to YouTube, disclosed on TikTok, and captioned differently on LinkedIn from the same source asset.
The safeguard is the human checkpoint. Under Autopilot, Kompozy runs a per-post review pipeline — the same review step that keeps invented statistics and off-brand language out of copy is the natural place to confirm that anything synthetic carries the disclosure a given platform expects before it goes live. Kompozy is not a provenance tool and will not sign a C2PA credential or read a SynthID signal for you; its role is upstream of detection, at the publishing decision. And it also produces net-new formats a single image generator cannot — persona and avatar video, carousels, blogs, newsletters — where the same disclose-at-publish discipline applies. The point that outlasts this specific announcement: once the automatic label is gone, the durable advantage is a publishing process where a person decides, per platform, that the content is on-brand, worth reading, and honestly labeled — not a corner sparkle doing that job by accident. The failure mode a watermark could never fix is AI slop; the one it was accidentally covering is undisclosed publishing, and both are solved at the same review step.
A visible AI watermark is an on-image mark — like the small Gemini sparkle in the corner, or a "Made with AI" badge — that a generator stamps onto its output so anyone looking at the file can tell it was AI-made. It is a disclosure aimed at humans, distinct from an invisible watermark like SynthID, which is a machine-readable signal embedded in the pixels or audio that a person cannot see but a detector can read.
No. On August 14, 2026, Google made only the visible watermark optional through a Settings toggle. The invisible SynthID watermark and C2PA provenance metadata stay embedded in every image, video, and song Gemini generates regardless of the setting. Google's framing was that it is balancing creative control with safety — the visible label is now your choice, but the origin of the file remains detectable through the invisible layer and tools like Gemini or Search.
No. Turning off the visible mark removes only the badge a human can see. SynthID and C2PA metadata remain in the file, so a platform or a detection tool can still identify it as AI-generated. Removing the visible watermark changes what your audience sees, not whether the content can be traced to an AI origin. Treat a watermark-free export as un-labeled, not un-detectable.
Yes, wherever a platform or a law requires it. The visible watermark was never your legal disclosure — it was the generator's. Platforms like YouTube, TikTok, and Meta have their own AI-disclosure rules and labeling toggles, and the EU AI Act requires marking synthetic media, all of which apply to your published post independent of whether a corner badge is present. Removing the badge shifts the disclosure job onto you, at publish time.
They answer different needs. The visible mark is a branding-and-trust signal aimed at viewers, and creators objected that it cluttered otherwise-usable output and made Gemini content look second-class next to competitors. The invisible SynthID and C2PA layer is the safety-and-provenance mechanism: it survives cropping and re-encoding, cannot be casually removed, and lets platforms and Google's own tools verify AI origin. Google kept the layer that does the real provenance work and made optional the one that was mostly cosmetic.
Kompozy is an AI content generation and multi-platform publishing engine, and disclosure lives at its publish layer — exactly where the responsibility landed once the visible watermark became optional. From one source it produces posts, images, carousels, blogs, newsletters, and persona or avatar video, then schedules and fans them across eight social platforms plus blog and email. Because each platform is published individually, you set the caption and any AI label per destination, and the per-post review step under Autopilot is a human checkpoint where you confirm the disclosure before the post ships — so a clean, watermark-free asset still goes out correctly labeled on every platform that asks for it.
A visible AI watermark is the on-file mark — like Gemini's corner sparkle — that tells anyone looking that content is AI-generated. On August 14, 2026, Google made Gemini's visible watermark optional through a Settings toggle, while keeping the invisible SynthID watermark and C2PA metadata embedded in every image, video, and song. The label you could see is now optional; the provenance you cannot see stays. For creators, that shifts AI disclosure from automatic to a per-platform decision you make at publish time.
Get started → · ← All guides · Compare Kompozy vs other tools