Google allows visible Gemini watermarks to be removed

Google shifts the balance between visible disclosure and technical traceability

Google now allows users of its Gemini tools to disable the visible watermark applied to certain creations generated by artificial intelligence. The information was reported by The Verge, in an article titled “You can now turn off Google Gemini’s visible watermarks”. The change concerns content produced by the group’s generation tools, including images, videos and audio productions.

This development does not mean, according to the information reported by the US media outlet, that all identification mechanisms have been abandoned. Google’s invisible SynthID marking remains in place. The decision therefore consists of separating two functions that have until now often been conflated in public debate: on the one hand, a signal directly perceptible to a person viewing or using content; on the other, technical information embedded in the file and intended to help detect its synthetic origin.

The visible watermark serves a purpose of immediate transparency. When displayed on an image, video or other creation, it indicates without any special tool that the content comes from a generation system. It is easy to understand and requires neither technical knowledge nor a specialized platform. But it also changes the appearance of the result. For users seeking to produce a visual, clip or audio element incorporated into a broader creation, this mark may be perceived as an aesthetic or operational constraint.

SynthID serves a different purpose. Presented by Google DeepMind as a digital watermarking technology, the system aims to insert a signal that is not necessarily visible or audible during normal use. Google has gradually associated SynthID with several types of content generated by its models, including images, audio, text and video. The aim is to maintain a technical verification capability without necessarily placing a label or logo in the final output.

The change seen in Gemini is therefore less a removal of traceability than a shift in transparency. People receiving content will not automatically benefit, in every case, from a visual clue to its generated nature. However, a detection mechanism is still intended to accompany the file. This distinction is central at a time when generation tools are becoming capable of creating material that is increasingly difficult to distinguish from human productions or traditional recordings.

The decision also comes in a competitive context. Publishers of models and creative software are simultaneously seeking to attract creators, reassure businesses and address regulators’ concerns. Users are asking for usable, clean results that can be easily integrated into production workflows. Authorities, media outlets, platforms and researchers, meanwhile, stress the need to be able to identify synthetic content, particularly when it can be mistaken for authentic documents.

Google is not abandoning one of these objectives in favor of the other: the group is changing how they coexist. This is precisely what makes the announcement significant for AI regulation. It illustrates the difficulty of defining what “good” marking is. Should it be directly visible to every recipient? Should it be robust against file transformations? Should it remain unobtrusive so as not to diminish the perceived quality of a creation? Or should it combine several mechanisms, depending on the risk associated with the content and its distribution context?

From Bard to Gemini, the question of content origin has become central with multimodal tools

The issue has not emerged in isolation. Google first launched Bard, before renaming its assistant Gemini in 2024 and extending that brand to a broader family of AI models and experiences. This development accompanied a more general movement in the industry: conversational assistants are no longer limited to text. They are expected to understand and produce images, sound, video, code and documents combining several formats.

This rise of multimodality has transformed the question of provenance. Generated text can be copied, modified or summarized with relative ease; identifying its source is already complex. For an image, voice or video, the stakes are of a different nature. These formats are regularly interpreted as traces of reality. They can be used to illustrate information, promote a product, represent a person or document an event. Their potential for persuasion and confusion is therefore greater.

Google DeepMind presented SynthID in 2023 as a technical response to this issue for AI-generated images. The principle is based on embedding a digital signal in the content produced, with the aim of making it detectable while limiting its impact on perception of the result. Google subsequently communicated about extending this approach to other formats. Keeping SynthID after making it possible to disable the visible watermark is part of this continuity: the group retains the technical layer it developed to associate certain content with its generation systems.

However, a marking technology must be distinguished from an absolute guarantee of provenance. An invisible watermark, however sophisticated it may be, is useful only if it can be sought, detected and interpreted by the relevant parties. This requires verification tools, suitable interfaces and, above all, an understanding of what a positive or negative result means. Detection of a SynthID signature may indicate that content was produced or modified by a relevant Google tool. Its absence does not, on its own, make it possible to conclude that content is authentic or that it was not generated by another system.

This asymmetry is important. The generative AI ecosystem is not centralized around a single provider. Open models, commercial services, image-editing software and editing tools may successively be involved in the same file. An image may be created in one tool, transformed in another, cropped, compressed and then published on a platform that does not retain all the technical information associated with the original document. The persistence of a mark then depends on the nature of the signal, the transformations applied and compatibility between systems.

Using a visible watermark addressed this difficulty through a very direct method: the signal did not depend on a detection operation. It could be seen immediately, provided it was not cropped or removed during editing. But its conspicuous nature could also encourage users to seek unmarked results or turn to competing tools. Google now appears to acknowledge, through this disabling option reported by The Verge, that acceptance of generated content also depends on its usability in creative contexts.

The debate extends beyond Gemini alone. Adobe has strongly highlighted its Content Credentials, based on provenance information linked to files. Microsoft, OpenAI, Meta and other companies have also participated, to varying degrees, in initiatives or mechanisms concerning the transparency of generated content. The Coalition for Content Provenance and Authenticity, or C2PA, is developing a technical standard aimed at documenting the origin and modification history of certain digital content.

These approaches do not overlap exactly. A watermark may seek to survive file transformations. Provenance metadata may document a creation chain, but it may be removed during export or publication. A visible label is accessible to the public, but it may degrade the output or be erased. Google’s decision therefore does not settle a technological debate; on the contrary, it confirms that no single method addresses all uses, all formats and all forms of manipulation.

The visible watermark creates a quality issue, but its removal changes the public’s experience

For creators, the ability to remove a visible watermark can have immediate value. In advertising, graphic design, illustration, video previsualization, product prototyping or the preparation of mockups, a mark placed in the image can prevent the output from being used as is. It can be particularly inconvenient when content is intended to be inserted into a presentation, publication or broader composition. The issue is not merely aesthetic: it also concerns visual consistency, framing and readability.

The same observation applies to the video and audio formats mentioned by The Verge. A marker overlaid on the image can remain visible throughout a sequence. An audible signal, if such a mechanism is used in a given context, can affect listening. Yet generative tools are often used in iterative production processes, in which the final content is edited, combined with other elements and adapted to a distribution format. A visible mark can therefore become an obstacle, including in uses that are in no way deceptive.

The option offered by Google therefore responds to users’ demand for control. It implicitly recognizes that marking should not necessarily be equated with a visual penalty applied to all creations. Under this interpretation, SynthID becomes the layer of technical accountability, while the user can decide the final appearance. This is a concept closer to certain provenance mechanisms embedded in files than to systematically displayed warnings.

But this gain for the creator results in a loss of visibility for the person receiving the content. An internet user faced with an image without a visible watermark will not know, simply by looking at it, that it was produced by Gemini. They will need a tool compatible with SynthID detection, know how to use it, and have access to the file under conditions that allow this verification. In practice, this step is more demanding than reading a warning placed on the content.

The difference is particularly noticeable in fast-sharing environments: messaging services, social networks, screenshots, reposts, compilations and successive exports. A reader, viewer or listener does not always control the original file. They may only have access to a recompressed, cropped or edited version. The technical possibility of detecting a signature therefore does not automatically translate into effective transparency when the content is viewed.

This tension explains why debates on synthetic content sometimes contrast two distinct concepts: traceability and labeling. Traceability seeks to preserve verifiable information on the origin or generation of content. Labeling aims to inform the public directly, in the context in which content is seen or heard. The former can be discreet, automatable and suited to after-the-fact verification. The latter is more explicit, but it depends on interface choices and may be challenged by creators who see it as limiting the use of their productions.

The Gemini case shows that a company can choose to preserve the former while making the latter optional. This choice does not say that the visible watermark is useless. It says that it is no longer treated as the sole form of transparency. The question now will be which parties can access SynthID information, under what conditions and with what level of trust. For journalists, platforms, researchers or verification teams, the robustness and accessibility of detection will be decisive.

A too-broad interpretation of the change should also be avoided. Disabling a visible watermark on Gemini creations does not mean that all AI-created content can become anonymous, nor that Google is abandoning detection of its own content. The source cited in the brief explicitly states that SynthID is being retained. This is therefore not the disappearance of all technical information, but a change in how that information is made perceptible.

According to The Verge, Google allows the visible watermark to be removed from Gemini creations while retaining the invisible SynthID marking.

A decision that resonates with European transparency obligations

For Europe and France, this development must be read in light of the European regulation on artificial intelligence, commonly called the AI Act. The text provides for transparency obligations for certain systems and certain uses of synthetic content. It gives significant importance to the ability to detect that content has been generated or manipulated by AI, particularly in cases where the appearance of the content may create confusion with a real situation.

The regulation is not limited to a visual watermark requirement. Its approach is broader and relies in particular on information in a machine-readable format, as well as on the detectability of synthetic content in certain circumstances. This distinction is essential in analyzing Google’s decision: retaining SynthID potentially fits within a technical marking logic, while removing the visible signal is more a matter of product and user-experience choice.

However, regulatory compliance and public understanding are not synonymous. A machine-readable mechanism can meet a traceability objective without allowing everyone to know at first glance that content is generated. Conversely, a visible notice may be immediately understandable but provide no reliable information about the tool used, the file’s history or the changes that followed its creation. The two levels can be complementary rather than interchangeable.

In France, this distinction will be of direct interest to companies that use generative tools in their communications. A brand, agency, publisher, newsroom or local authority may use AI to illustrate a campaign, create an explanatory video, generate audio branding or produce internal materials. The ability to remove a visible mark makes use easier within a demanding visual identity. But it also increases the organization’s editorial responsibility when it decides whether, how and when to inform its public about the involvement of a generative system.

The law does not cover all situations in the same way. The level of risk is clearly not the same between an abstract illustration designed for presentation material and a realistic video likely to be interpreted as testimony of an event. The expected transparency also varies depending on the context: commercial communication, information, artistic creation, training, simulation, entertainment or interaction with a digital service. A general option to remove a visible watermark therefore does not exempt users from assessing the actual use of the content.

The debate also concerns platforms. They play a concrete role in displaying, retaining or removing information associated with files. An invisible mark that is not exposed in a consumer interface may remain inaccessible to most users. Conversely, a platform may choose to add its own label when it detects that content comes from an AI tool, if it has an exploitable signal or a declaration from the author. The value of marking thus depends on an entire chain: model provider, creator, editing software, file format, platform and viewing interface.

In the French-speaking market, where language and local cultural contexts are often less well covered by international automated moderation mechanisms, this interface issue takes on a particular dimension. Technical provenance information has a collective effect only if it can be used by those who verify content in French: newsrooms, regional media, fact-checking organizations, institutions, communications departments and platforms. The presence of a signal in the file does not, on its own, resolve the difficulties of contextualization, source verification and media literacy.

Finally, Google’s decision highlights a point that is often overlooked: identifying AI does not automatically make it possible to determine whether content is reliable. A generated visual can be clearly labeled and still be used in a legitimate, creative or educational way. A real photograph may contain no trace of AI and be taken out of context. Provenance is important information, but it is only one element of a broader verification process, which includes the date, location, source, framing, any alterations and the intent behind distribution.

The market will have to choose between interoperable standards and fragmented transparency

The ability to remove Gemini’s visible watermark reinforces the need for verification tools capable of working beyond a single interface. If SynthID remains attached to the relevant creations, the practical question is its recognition in an environment with many providers. A signature specific to one actor is useful for identifying its own content, but it does not, on its own, constitute a universal language of provenance.

This is where standardization efforts become important. Initiatives such as C2PA pursue the idea that content can carry information about its origin and transformations in a structured manner. Google, for its part, promotes SynthID as a watermarking mechanism embedded in content generated by its models. These two families of approaches do not necessarily address the same constraints: metadata and provenance information describe a history; the watermark seeks to leave an imprint in the content itself.

For users, the issue is less about knowing the name of every technology than about being able to verify information easily. A newsroom that receives a video, a platform analyzing an upload or a company approving communications material should ideally be able to access an understandable result: was the content generated or modified by AI, by which tool when that information is available, and has that information been retained? Without interoperability, each verification risks depending on separate tools, proprietary procedures and levels of reliability that are difficult to compare.

Google is making a choice here that gives greater priority to the creator’s control over the file’s appearance. This may encourage the adoption of generative tools in professional uses, where a visible watermark is sometimes incompatible with production-quality expectations. But the group will also have to demonstrate that SynthID offers a solution that is genuinely usable by third parties. The persistence of marking is not enough if detection is limited to restricted environments or if the public cannot understand the results obtained.

This situation could reinforce a division of roles. Creators would have access to content without mandatory visual indication. Platforms and verification services would use technical signals to classify, contextualize or label files. The public, in turn, might see information produced by the distribution interface rather than a mark directly embedded in the content. This model has the advantage of preserving the visual output, but it shifts a significant part of transparency onto digital intermediaries.

It also has limitations. Information added by a platform may disappear when a file is downloaded and then republished elsewhere. Content can circulate outside major platforms, in messaging services, on an independent website or in a shared document. Actors seeking to deceive an audience may try to avoid channels that display warnings. No marking solution therefore eliminates the need for distribution policies, human verification and user training.

In the long term, the announcement reported by The Verge could serve as a test for the industry. If retaining SynthID allows the relevant services to effectively detect Gemini content despite the absence of an apparent watermark, other providers could be encouraged to likewise separate the output’s aesthetics from provenance information. If, on the contrary, removing the visible signal makes identification much less accessible in ordinary use, pressure could increase in favor of more explicit labels for the most realistic or most sensitive content.

For French and European stakeholders, the issue will not be only whether a file carries a signal. It will be determining who can read it, at what point, with what compatibility between tools and under what editorial responsibility. Google’s choice places SynthID at the heart of this question: transparency for AI-generated content may become less and less visible in the work itself, and increasingly dependent on the technical infrastructure surrounding it.