New Media Watermark Setting in Gemini
Google now provides a direct toggle to remove visible watermarks from AI-generated images, videos, and music produced in Gemini. The setting appears under the label “Media watermark” and applies across Gemini as well as the company’s separate AI video generator, Flow. When users turn the option off, the sparkle icon that previously appeared in the bottom-right corner of outputs disappears.
The change covers content created with the Nano Banana and Omni models. Visible marks are no longer required, yet Google continues to embed invisible SynthID watermarks and C2PA metadata in the files. Josh Woodward, vice president of Google Labs, Gemini, and AI Studio, confirmed that these background signals remain active so that verification tools can still detect the origin of a file.
Users retain the ability to check whether material carries these markers by querying Gemini or Google Search directly. The adjustment therefore separates the choice of a visible indicator from the technical record that supports later authentication. Details on rollout timing across additional surfaces remain limited in the initial announcement.
How to Toggle Visible Watermarks Off
Google directs users to disable visible watermarks through a single setting inside the Gemini app. The option sits in the settings menu under the label “Media watermark.” Users reach it by navigating to Settings and then selecting Media watermark.
This control will appear in the coming days for both the Gemini app and the Google Flow apps. Once enabled, the toggle removes the visible sparkle overlays from AI-generated images, video, and audio. Invisible SynthID and C2PA watermarks remain embedded in all output regardless of the setting.
Current announcements provide no further details on confirmation dialogs, default states, or platform-specific differences. Details on exact behavior after the toggle remain limited in the initial reports. The change follows user reports that many already manually edit visible marks out of generated files.
Google positions the menu item as the direct response to those preferences while preserving the invisible detection layer. Users seeking to verify AI origin after the change can still query Gemini directly for confirmation on specific content.
Invisible SynthID and C2PA Still Applied
Google confirmed that its decision to make visible watermarks optional does not extend to the underlying detection systems. Invisible SynthID watermarks and C2PA metadata remain embedded in all AI-generated images, videos, and audio created with Gemini, regardless of user settings.
The company announced on Friday that these technical markers continue to support verification even after the visible labels can be turned off. Users who want to check whether a file was produced by the model can still query Gemini directly or run a search to surface the embedded signals.
Josh Woodward, VP for Google Gemini, described the approach as a balance between creative control and safety. The quote noted that invisible SynthID watermarks and C2PA metadata stay active so that Gemini or Search can still identify AI-generated content.
The visible watermark toggle will appear in Settings > Media watermark inside the Gemini app and will reach Google Flow in the coming days. Because the remaining markers are not visible on the surface, they require an active check rather than casual inspection.
This setup preserves a verification path for platforms and users who perform the extra step, while removing the automatic on-screen indicator that previously appeared on every output.
Verification Methods Remain Available
Google made the visible watermark optional for generations from its Nano Banana, Omni, and Lyria models, yet the company kept two other identification systems intact. The invisible SynthID watermark and C2PA standard-related metadata continue to attach to images, video, and audio regardless of the new setting.
Josh Woodward, vice president for Gemini, stated that the change lets users retain creative control while preserving detection capabilities. He noted that people can still query Gemini or Search to determine whether a given image originated from an AI model. The toggle appears under Settings > Media Watermark in Gemini and in the Flow video editor, with Search support scheduled to follow.
Developers receive an additional tool. Google is releasing an open-source library named Credentio that allows applications to embed local validation checks without relying on external services. The feature begins rolling out within days.
The approach separates visible presentation from persistent verification. Invisible markers and metadata remain embedded even when users disable the on-screen mark, and the same query interfaces that already support detection continue to function.
Comparison to OpenAI, Meta, and Anthropic
Google's decision to offer a Settings > Media Watermark toggle for visible marks on Gemini, Flow, and eventually Search content stands in contrast to the approach taken by Anthropic. The research material notes that this Google update follows Anthropic's inclusion of watermarks in text and files generated by Claude, a step taken to meet EU regulatory requirements. That move by Anthropic drew debate because it applied watermarks without an apparent user-facing option to disable them.
No details appear in the source material on OpenAI or Meta watermark policies. Information on whether those companies embed visible marks, hidden signals such as SynthID, or C2PA metadata, or whether they provide toggles, remains outside the scope of the provided reporting.
Google's implementation keeps SynthID and C2PA intact even after the visible toggle is turned off. This preserves detection capabilities for downstream tools while giving users control over the on-screen Sparkle element. The distinction matters for developers building moderation or provenance systems, since the hidden layer continues to function across resizing, compression, and other edits. Further announcements from other providers would be needed to complete a full side-by-side view.
User Workarounds Before the Update
Prior to the toggle for disabling visible sparkle watermarks, the research material provides no concrete examples of user techniques to strip or bypass those marks on Gemini outputs. Instead, it highlights the role of SynthID as the persistent layer that remains regardless of visible changes.
SynthID is designed to survive resizing, recoloring, and compression. This allows Gemini or Search to still detect the content as AI-generated even without the visible mark. The same invisible watermark applies by default to every generation from Imagen 4, including the Ultra variant, through both Vertex AI and the Gemini API. It also covers Gemini 2.5 Flash Image, Gemini 3 Pro Image, and every frame produced by Veo video generation.
Verification APIs exist for teams that need to check provenance after the fact. The source notes that daily.dev offers a channel for tracking updates to these detection methods. No other providers appear in the supplied tables with alternative approaches for Google models.
Details on practical steps users took to remove visible marks before the toggle are still emerging from official documentation. The documented priority remains the non-visible SynthID layer that operates independently of any surface-level edits.
Impact on Quick AI Content Detection
Google's decision to let users disable visible sparkle watermarks shifts reliance for rapid identification of AI-generated content. The visible markers previously allowed casual observers to spot Gemini outputs at a glance across images, video, and audio. With the toggle enabled, that immediate visual cue disappears for those who choose to remove it.
Non-visible SynthID watermarks remain embedded by default in every output from Imagen 4, Imagen 4 Ultra, the broader Imagen family through the Gemini API, Gemini 2.5 Flash Image, and Gemini 3 Pro Image. Google documentation states that all generated images continue to carry these invisible markers, and the company provides a verification API to read them. The Gemini app itself can scan for SynthID to confirm whether an image originated from or was edited by Google models, though this detection works reliably only on Google-generated content.
Casual users lose the fastest method of spotting AI material when the sparkle is turned off. Detection now requires either the dedicated app scan or access to the SynthID verification tools, which demand more than a quick look. The change does not eliminate watermarking but moves the process from visible and immediate to technical and deliberate.
Google Labs VP Statement on Balance
No statement from a Google Labs vice president on the balance between visible watermarks and user control appears in the available material. The documentation centers on third party tools that strip overlays rather than on internal product decisions at Google.
The provided details describe an extension that removes visible Gemini watermarks, visible Nano Banana watermarks, and visible sparkle or diamond badges from content generated in Gemini and Google AI Studio. It leaves SynthID untouched because that system embeds invisible provenance information. The same holds for stock photo watermarks, photographer credits, and marks from unrelated platforms.
Processing occurs locally inside the browser with no files uploaded to external servers. The tool handles still images and processes video content frame by frame. It supports Gemini, Nano Banana, Nano Banana Pro, and Google AI Studio outputs. Questions about whether the cleaning step reduces image or video quality receive no answer in the source.
The material confirms that AI generated media can remain detectable through SynthID even after visible elements disappear. No executive commentary explains why Google might add an official toggle for visible markers or how such a feature would interact with the persistent invisible layer. Details on this are still emerging.
Rollout Timeline for Gemini and Flow Apps
No specific schedule appears in the available documentation for when the toggle to disable visible sparkle watermarks will reach Gemini or Flow applications. The research material focuses on SynthID detection tools rather than product release plans, leaving concrete dates and phased deployment steps unaddressed.
What is documented centers on the underlying SynthID system itself. Google DeepMind introduced SynthID as an invisible watermarking method for Gemini and Imagen outputs, later extended to text and audio formats. The material notes adoption by OpenAI for ChatGPT image generation, with SynthID applied to DALL·E images since May 2026 according to the detector comparison table. This table also records that the watermark survives JPEG compression, resizing, cropping, and color adjustments for images, while text versions rely on token probability shifts that persist through paraphrasing.
The absence of rollout information for the visible watermark toggle means any assumptions about integration with browser-based local processing extensions or broader app updates would exceed what the sources confirm. Details on this are still emerging. The material instead emphasizes functional limits of existing detectors, such as their inability to handle stock photo watermarks or non-SynthID content from Mid, Stable Diffusion, or Claude.
Users seeking to manage watermarks in Gemini-generated media therefore have no verified timeline to reference for the new toggle feature. Further official announcements from Google would be required to clarify deployment across apps.
Broader Shift in AI Watermark Practices
Google's decision to let users disable visible sparkle watermarks on Gemini outputs fits into a wider pattern where providers move away from overt markers toward embedded, invisible systems. SynthID, developed by Google DeepMind, inserts signatures directly into AI-generated images, audio, and text at creation time. These signatures remain undetectable to the eye or ear yet persist through screenshots and compression.
Multiple services now rely on comparable approaches. Google Gemini applies SynthID to text, images, and audio. Google Imagen uses it for images. OpenAI incorporates it into DALL·E images and plans expansion to text. The OpenAI API and Google Cloud AI follow the same pattern. Claude from Anthropic and tools such as Mid and Stable Diffusion appear on detector lists that target SynthID content across providers.
Detection services report average results under three seconds with accuracy above 99 percent on watermarked material. At the same time, third-party offerings have emerged that strip or bypass these signatures. One service, Rephrasy.ai, advertises the ability to humanize AI text so it reads as original and evades SynthID detection.
The combination of widespread invisible watermark adoption and active removal tools shows that the field continues to adjust its technical and policy choices rather than settle on a single standard.