The rapid spread of generative AI tools has made it routine to encounter images, videos, and audio whose origins are no longer obvious at a glance.
Distinguishing camera-captured material from synthetic versions has become a practical problem for journalists, platforms, and ordinary users alike, prompting a range of technical responses that aim to leave a trace of how a piece of media was made or altered.
Google has spent several years building one of those responses through SynthID, an invisible watermarking system that embeds signals into content produced by its own models.
The technology has been applied to images, video, and audio at large scale, and verification capabilities have gradually moved from a dedicated portal into the Gemini app, Search, and Chrome.
Alongside it, the company has adopted the C2PA standard for Content Credentials, which records provenance information such as capture device, editing tools, and whether generative systems were involved.
By mid-2026 those credentials were being added at the point of creation on Pixel devices and across various Google products, with verification features rolling out in Gemini and planned for broader surfaces.
On the front-end, Google has that visible layer of labeling applied by default.
Users using Gemini to generate images will see that small Gemini logo in the corner of the image. As for videos, they have overlays, and music tracks displayed a corresponding indicator.
User feedback, gathered in public threads and ranked among other product requests, placed the removal of those visible marks high on the list of requested changes.
Google finally addresses this.
The company responded by introducing a setting that lets people turn the visible watermark on or off inside Gemini and the Flow video tool, with Search support expected later.
The option covers output from the models used for images, video, and music, though it remains unavailable in jurisdictions where law requires the marks to stay visible, and it is restricted for certain account types.
The invisible layers are unaffected.
SynthID signals and C2PA metadata continue to be written into every generated file, so a user can still ask Gemini or Search whether a given piece of media carries those markers.
The change therefore separates the human-readable indicator from the machine-readable record, treating the former as a matter of creative preference while preserving the latter for verification.
At the same time Google said that it has released Credentio, an open-source C++ library that implements local validation of C2PA Content Credentials.
The library is the same code already running inside dozens of the company's own products. It operates entirely on-device, parses manifests and signatures without sending files to external servers, and returns detailed integrity results.
Developers can link it into desktop, mobile, or server-side applications that need to check provenance at scale while keeping media private and latency low. Support currently covers C2PA specification versions 2.2 and 2.4, with plans to expand into credential generation as the standard evolves.
Taken together, the optional visible mark, the persistent invisible systems, and the newly shared validation library reflect an incremental adjustment rather than a reversal.
The technical means of establishing origin remain in place even as the surface-level presentation becomes configurable.
