A research page describing an invisible signal
One entry here is read from a research page. It is the fullest description of a mark on this site, and the only one that states which transformations the mark is designed to survive. As of 2026-09-22.
| Entry | What the page describes | What the register does with it |
|---|---|---|
| Google (SynthID) | Scope, timing, robustness and two checking routes | An invisible mark, with no tier and nothing offered as removable |
Inclusion rule. Entries whose recorded values come from a page describing how a marking system works, published by the organisation that builds it rather than by a seller of access to it. Order. Alphabetical by the name in the first column.
1Four things stated that no commercial page states
Google's SynthID page says where the watermark is embedded, that it is added the moment content is created, which transformations it is designed to stand up to, and how a file can be checked. Cropping, added filters, changed frame rates and lossy compression are named individually. Nothing else in this register names a single transformation, let alone four.
Those four statements happen to be exactly what a receiving party needs. A publisher handed a file wants to know whether a signal is present, whether ordinary handling destroyed it, and who can read it. A page written to explain a mechanism answers all three; a page written to sell a plan answers none of them.
2The genre inverts the register's own columns
This entry is kept under implementers rather than in the generator table because the four columns assume a shape it does not have. Nothing is visible, so the visible-mark columns have no content. No tier lifts the signal, so the crossing point is not merely unstated but inapplicable. Nothing is offered as removable, so recording a no in a removal column would be reading a business decision into an engineering one.
Its generator-side counterpart is on a different page: Veo's documentation names the same scheme and the platform that verifies it. Keeping the scheme and the generator as separate rows is deliberate, because a scheme can be described in full while the products using it stay silent, and that is what the rest of this register shows.
3What a research page cannot be read as
A description of a mechanism is not a statement about any particular file. The page speaks about generative products across a company's range, and the register does not turn that into per-product cells. Where a generator's own documentation names the scheme, that generator gets a filled cell; where it does not, the scheme's existence changes nothing on its row.
The other limit is testing. Designed to stand up to a transformation is the vendor's own phrasing, recorded as such, and this site produces nothing by running files through an editor to see what happens. A claim about robustness and a measurement of robustness are different objects, and only the first one is written down anywhere public.
4Sources
- Machine-readable markingAn invisible watermark is embedded across Google's generative AI consumer productsproduct-wide
- When it is writtenThe watermark is added the moment content is created and does not change image or video qualityat generation
- RobustnessDesigned to stand up to cropping, added filters, changed frame rates and lossy compressionas stated by the vendor
- DetectionGemini checks an uploaded image, video or audio clip for the watermark, and a separate SynthID Detector portal is in testing with journaliststwo published routes
Every line above is read from the document named beside its row, on 2026-09-22. The same values in their register form are on the support table, and the rules for what a cell may hold are on the what earns a row page. Other kinds of source document: plan grids, developer references.