GateTrue

What each generator documents about marking its output

SynthID: what Google publishes about its invisible watermark

SynthID embeds an invisible watermark across Google's generative AI consumer products. Google states it is added the moment content is created, does not change quality, and is designed to survive cropping, filters, frame-rate changes and lossy compression. As of 2026-09-12.

Where an invisible watermark sits relative to the other two marking layersThree layers can carry a mark. A visible watermark is drawn into the picture and can be seen and cropped. An invisible watermark is embedded in the signal itself at the moment of creation and is stated to survive cropping, filters, frame-rate changes and lossy compression. Container metadata sits outside the picture entirely and does not survive a re-export.Outermost firstVisible watermarkDrawn into the frame. Anyone can see it and anyone can crop it.Invisible watermarkEmbedded in the signal as the content is created. Stated tosurvive cropping, filters, frame-rate changes and lossy…Container metadataOutside the picture. Removed by a re-export through an ordinaryeditor.Only the middle layer is claimed to survive an export
Fig. 1 The three layers fail in different ways, which is why this entry sits in its own column rather than alongside the visible marks.
The Google SynthID page as it appeared on 16 September 2026.
Fig. 2 One of three entries here describing an invisible mark, with Veo and Amazon Nova Reel, and the only one read from a research page.Screenshot of deepmind.google/science/synthid, captured 2026-09-16. Reproduced to show where this entry was read from.
  • Machine-readable marking
    An invisible watermark is embedded across Google's generative AI consumer productsproduct-wideGoogle DeepMind, SynthID / recorded 2026-09-12
  • When it is written
    The watermark is added the moment content is created and does not change image or video qualityat generationGoogle DeepMind, SynthID / recorded 2026-09-12
  • Robustness
    Designed to stand up to cropping, added filters, changed frame rates and lossy compressionas stated by the vendorGoogle DeepMind, SynthID / recorded 2026-09-12
  • Detection
    Gemini checks an uploaded image, video or audio clip for the watermark, and a separate SynthID Detector portal is in testing with journaliststwo published routesGoogle DeepMind, SynthID / recorded 2026-09-12

1Why this sits in a different column from a visible watermark

Most generators recorded on this site document a mark a viewer can see, and tie its removal to a plan. SynthID is described the other way round: nothing is visible, nothing is offered as removable, and the stated design goal is that the signal outlives ordinary handling of the file. Those are different products serving different purposes, and putting them in one yes-or-no column would lose the distinction.

The robustness claim is the part most worth reading precisely. Google names four transformations the watermark is designed to stand up to: cropping, added filters, changed frame rates and lossy compression. Designed to stand up to is the vendor's phrasing and is recorded as such; it is not a guarantee, and no figure is published for how much of each transformation the signal tolerates.

2How Google says it can be checked

Two routes are published. Gemini will check an uploaded image, video or audio clip and report whether it finds a SynthID watermark. A separate verification portal, the SynthID Detector, is described as in testing with journalists and media professionals, with a waitlist rather than open access. A publisher who needs to verify at volume therefore has one route available today and one that is gated.

For anyone reading this site to answer a compliance question, the useful consequence is that detection and marking are separate capabilities. A generator can write a signal nobody outside the vendor can read, and a platform can read a signal that most generators never write.

3What the delivered file carries

Next to the marking question in “SynthID”, SceneMixer's pages describe its output like this:

You can recomposite with different end-card text, and in step-by-step mode you can reshoot individual segments. The finished file can also be downloaded and shared with a link.

SceneMixer, How to Make an Ad from One Photo — “Can I change the ad after it is finished?” (read 2026-10-13)

Dubbing replaces the audio of a finished clip; the picture was generated for a different line, so mouth shapes, pauses and shot length all belong to the original language. Native-language generation gives the video model the line in the target language before it renders the shot, so the performance, the pacing and the cut points are built for that language.

SceneMixer, Native-Language Dialogue vs. Dubbing in AI Video (2026): Why the Cast Should Speak the Line (read 2026-10-13)

4Sources

Read from Google DeepMind's SynthID page, checked 2026-09-12. How the fields on this site are defined is on the protocol page.