You can now turn off Google Gemini's visible AI watermarks. Here is what that actually changes

Google has quietly added a new toggle to its Gemini app and Flow video generator that lets users remove the small visible watermark stamped on AI-generated images, videos and music. The change, confirmed by the company, marks a shift in how one of the world's largest AI developers is approaching the increasingly contentious question of labeling synthetic media, and it arrives at a moment when regulators and platforms are moving in the opposite direction, pushing for more transparency around AI-generated content rather than less.

The visible watermark in question is a small, semi-transparent "sparkle" icon that Google's Nano Banana image model and Omni video and music model have stamped in the corner of their outputs since launch. It functions similarly to watermarks long used by other AI image generators: a quick visual cue that tells anyone looking at the content, without needing to investigate further, that it was produced by an AI system rather than captured or created conventionally. With the new setting, users can turn that visible marker off, producing outputs that look identical to unmarked, conventionally created content at a glance.

Critically, disabling the visible watermark does not remove Google's other tracking mechanism, a technology called SynthID. Unlike the visible sparkle icon, SynthID embeds an imperceptible digital pattern directly into the pixel or audio data of AI-generated content, a signal that survives common editing operations like cropping, compression and format conversion far better than metadata alone typically does. Google, along with a growing number of AI companies, has positioned SynthID as a way to allow verification of AI-generated content even after it has been stripped of visible markers or shared across platforms that discard metadata.

Alongside SynthID, Google embeds C2PA metadata in the content it generates — a technical standard, backed by a coalition of major technology and media companies, that records information about how a piece of digital content was created and edited. Unlike SynthID's embedded signal, C2PA metadata can be stripped relatively easily by re-encoding or re-uploading content through platforms that do not preserve it, meaning it offers weaker guarantees of persistence than the watermark embedded directly in the media itself.

Google's rationale for allowing users to disable the visible marker, according to the company, centers on legitimate creative use cases where a visible watermark interferes with a finished product — a filmmaker incorporating AI-generated b-roll into a broader project, for instance, or a designer using AI-generated elements within a larger composited image where a visible sparkle icon would be a distracting artifact. The company has framed the invisible SynthID layer as the primary mechanism for accountability, arguing that removing a cosmetic visual marker does not meaningfully undermine the ability to verify AI provenance when needed.

Critics of the change argue that visible watermarks and invisible provenance signals serve fundamentally different social functions, even if both are technically present. A visible watermark provides immediate, universal signal to any viewer, requiring no special tools or technical knowledge to notice that content was AI-generated. An invisible signal like SynthID, by contrast, requires a viewer to actively run detection software or use a specific verification tool to confirm AI origin — a step that almost no casual viewer of an image or video circulating on social media will ever take. In practice, disabling the visible marker means the vast majority of people encountering the content will see nothing distinguishing it from conventionally created media.

The timing has drawn particular scrutiny given the broader regulatory trend. The European Union's AI Act includes transparency requirements around labeling AI-generated content, and several other jurisdictions have introduced or are considering similar disclosure requirements, particularly as concern grows about AI-generated misinformation, deepfakes and the general erosion of trust in visual and audio media. Policy analysts note that a default toward invisible-only labeling, even when technically compliant with a given jurisdiction's specific rules, runs somewhat counter to the broader spirit of transparency-focused AI regulation.

Google is not the first major AI company to face this tension between creative flexibility and content transparency. Several competing AI image and video generators have offered similar options to disable visible watermarks while retaining some form of embedded metadata, reflecting an industry-wide struggle to balance user demand for clean, professional-looking outputs against growing societal pressure for clear AI content labeling. The debate mirrors earlier tensions in digital media around metadata stripping and content provenance more broadly, now playing out specifically in the context of generative AI.

For ordinary users, the practical takeaway is straightforward but easy to overlook: the absence of a visible watermark on an image, video or piece of music no longer indicates the absence of AI involvement, if it ever reliably did. Detecting AI-generated content, going forward, increasingly requires either specialized verification tools capable of reading invisible watermarks like SynthID, or a broader skepticism toward assuming any digital media is unedited or conventionally created without explicit confirmation, a shift that reflects less a single Google product decision and more the trajectory of an entire technology now capable of producing convincingly realistic synthetic media at scale.