Google lets Gemini users hide visible AI watermarks, but invisible markers stay

Google is giving Gemini users control over whether AI-generated images, videos and music show a visible watermark, while keeping invisible identification methods in place. The change matters because it makes synthetic media less immediately obvious to viewers even as Google says its underlying provenance tools will remain embedded.

The update adds a new Media watermark setting in Gemini and Google’s Flow video tool. When turned off, the sparkle-style mark that has appeared in the corner of content made with Google’s Nano Banana and Omni models disappears, but SynthID and C2PA metadata continue to be attached in the background.

What Google changed in Gemini

Google is moving from mandatory visible branding to an optional one for some of its AI media tools. That means creators using Gemini and Flow can now choose whether exported content displays the small on-screen mark that previously identified it as AI-generated.

The company’s decision affects the most obvious visual cue viewers had that an image, video or piece of music came from an AI system. The corner watermark had been an easy way to spot content made with Google’s generative models, but it also became easy to remove through editing.

How the new Media watermark setting works

When the setting is disabled, Google removes the visible sparkle watermark from supported AI outputs. According to Google Labs leadership, the change applies to content created with Gemini and Flow, including media generated by Nano Banana and Omni models.

What does not change is the hidden layer of identification. Google says it will keep embedding SynthID, its invisible watermarking system, along with C2PA metadata, so the company can still help confirm whether a file was produced by AI.

Google has said that even if the visible marker is switched off, the background identification signals will remain. In practice, that means the company is trading a clear on-screen label for a less noticeable verification system that can still be checked later.

Why Google is making visible watermarks optional

Google appears to be responding to a simple reality: visible watermarks have not stopped people from altering synthetic content. Users with even basic editing tools can crop, blur, mask or otherwise remove a corner mark, which reduces its practical value as a safety feature.

At the same time, many rival AI products have never used visible watermarks at all. Google’s change brings its tools closer to the industry norm, while still preserving provenance signals that are harder to notice and harder to strip out at a glance.

What this means for viewers

The trade-off is straightforward. Content shared from Gemini may become harder to identify instantly as AI-made, especially after the visible watermark is turned off. For ordinary users scrolling through social feeds or messaging apps, the loss of that visual marker could make synthetic media blend in more easily.

Google argues that invisible markers and metadata can still support verification. In theory, someone can ask Gemini or Search whether an image was AI-generated, and the system can check for the relevant signals. In practice, that is less immediate than seeing a logo-like mark on the content itself.

How does Google’s approach compare with rivals?

Google’s approach now sits somewhere between visible labeling and fully invisible provenance. OpenAI also does not place obvious watermarks on its image outputs, relying instead on SynthID and C2PA to identify synthetic content. Meta, meanwhile, has promoted its own Content Seal standard.

The contrast shows a broader split in the AI industry over whether transparency should be visible to users in real time or handled through back-end verification. Google’s update suggests the company is increasingly leaning toward the latter model.

What are SynthID and C2PA?

SynthID is Google’s invisible watermarking system for AI content. It is designed to survive common edits and help identify content even when the visible cues are gone. C2PA, short for Coalition for Content Provenance and Authenticity, is a broader industry framework for attaching provenance information to digital media.

Together, those tools aim to create a chain of identification that travels with the file. That does not guarantee that every piece of manipulated media can be perfectly flagged, but it gives platforms, search tools and investigators a better chance of checking authenticity later.

Why invisible markers matter

Invisible markers matter because they are much harder to remove casually than a visible watermark. They are not intended to reassure a person looking at a single image in isolation; instead, they support downstream verification by platforms, services and forensic tools.

That said, hidden markers are not a complete solution. They depend on the systems that read them, and they do not offer the instant clarity that a visible label can provide. That is why the debate over AI transparency continues to intensify as generative tools become more realistic and more widely available.

How this fits into the wider deepfake problem

Google’s announcement arrives at a time when synthetic media is getting harder to distinguish from real photos, videos and audio. The pace of improvement in generative models has made the deepfake problem more urgent for platforms, publishers, election officials and ordinary users.

The company’s change reflects a wider tension in AI policy. On one hand, visible labels can warn viewers immediately. On the other, too much on-screen branding can be ignored, stripped away or viewed as a hindrance by creators who want cleaner-looking outputs.

What regulators are watching

AI labeling rules are becoming more prominent, especially in Europe, where transparency requirements have begun to take effect. That creates an uneven policy landscape: some markets may push for obvious labels, while others may allow more discretion as long as hidden provenance is preserved.

Google said the new toggle will not be available in countries where visible watermarking is required. That limitation is important because it shows the company is not abandoning regulation; it is tailoring the feature to local legal demands.

Where the new toggle will and won’t be available

Google plans to expand the watermark setting to Search, extending the change beyond content creation tools into a major discovery platform. But the company also made clear that the rollout will stop short in jurisdictions that mandate visible disclosure.

That creates a two-track model. In some places, users will have more control over how polished their AI-generated content looks. In others, legal rules will keep the label in place to preserve transparency.

What users should expect next

For creators, the immediate effect is more aesthetic flexibility. For everyone else, the change may make it marginally harder to tell whether something was generated by Gemini at a glance. Search and verification tools may become more important as the visual signal recedes.

For Google, the move is also a competitive adjustment. As generative AI products become more similar in output quality, provenance tools are becoming part of the product strategy as much as the safety strategy.

Timeline of Google’s watermark approach

Google has shifted from a visibly branded AI output model to one that puts more weight on hidden identification. The following timeline summarizes the key developments behind the latest change.

How worried should people be about the change?

People should be moderately concerned, but not alarmed, because Google is not eliminating provenance signals entirely. The risk is less about a total loss of transparency and more about the disappearance of the easiest cue that a piece of media came from an AI system.

That subtle shift matters. A visible watermark is immediate and universal. Invisible markers are technical and dependent on tools, policies and compliance. As a result, the update could widen the gap between what creators want and what casual viewers can quickly verify.

What the decision says about the AI industry

Google’s move shows that the AI sector is still experimenting with how to label synthetic content without making it unusable or unattractive. The industry is converging on invisible provenance systems, but not yet on a single standard for public-facing disclosure.

That leaves users in a complicated position. The content may look more natural and professional, but the responsibility for identification is shifting away from the image itself and toward the tools used to inspect it.

As AI-generated media becomes more convincing, those tools may become essential. Google’s update is a reminder that in the next phase of generative AI, the battle over authenticity will be fought as much in metadata as in pixels.

Key facts at a glance

• Google now lets Gemini and Flow users turn off a visible AI watermark.
• The visible sparkle mark applies to content made with Nano Banana and Omni models.
• Invisible SynthID and C2PA signals remain embedded in the media.
• Google plans to bring the setting to Search, but not where visible labeling is legally required.
• Other AI companies largely rely on invisible provenance rather than on-screen watermarks.