Which provenance standards get named, and by whom
Two schemes are named anywhere in this register. Both appear in technical documents, and no pricing page names either. Three generator rows and one implementer entry carry them. As of 2026-09-22.
| Entry | Machine-readable marking | Document read |
|---|---|---|
| Amazon Nova Reel | Invisible watermark plus C2PA Content Credentials | AI service card |
| Google (SynthID) | An invisible mark, written at generation | A research page |
| Sora | C2PA on every asset, standard named | Sora 2 system card |
| Veo | SynthID, with a verification platform named | Veo API documentation |
Inclusion rule. Entries whose own pages name a provenance scheme, a metadata standard or a described signal. A general mention of metadata without a name does not qualify. Order. Alphabetical by the name in the first column.
1Two schemes, and which layer each one occupies
One is a credential standard: a manifest attached to the container around the picture, which is what makes it readable by ordinary tools and removable by ordinary handling. Two generator rows name it, and one of them says the manifest carries the model and platform that produced the file.
The other is an in-picture signal, named by one generator row and described in full by the organisation that builds it. That entry publishes where the signal is embedded, that it is written the moment content is created, and which transformations it is designed to survive, which no generator documentation here does.
2No pricing page names either
Nine entries in this register are read from a pricing grid and not one of them names a scheme. That is structural rather than evasive: a provenance manifest is not a feature anybody upgrades to buy, so a document built around upgrades has nowhere to put it.
Which makes the naming of a standard a reliable signal about the genre of the source page rather than about the seriousness of the company. Every named scheme here comes from a system card, a service card, developer documentation or a research page, and the split holds across the whole register without exception.
3What naming a scheme does and does not settle
It hands a reader something to look up, which is the test a sentence has to pass before it fills the file-contents column. It does not say the signal is present in any particular file, and it does not say who can read it: one of these entries describes its detection tooling as internal to the vendor.
Nor does naming a standard imply durability. One vendor publishes the limit of its own mechanism, saying a check works unless the metadata has been removed. The others name a scheme and leave its fragility for the reader to work out from the specification.
4Sources
- Machine-readable markingIn the same list: C2PA metadata on all assets, providing verifiable origin through an industry standardstandard named
- Machine-readable markingVideos created by Veo are watermarked using SynthID, described as the vendor's tool for watermarking and identifying AI-generated contentstated under Limitations
- Machine-readable markingAmazon Nova Reel 1.1 adds Content Credentials to videos it generates, built on the C2PA specification, carrying the model and the platform used to generate the videostandard named
- Machine-readable markingAn invisible watermark is embedded across Google's generative AI consumer productsproduct-wide
Every row above is read from the vendor page named beside it, on 2026-09-22. The same values in register form are on the support table, the rule for this column is on the machine-readable marking column page, and what may fill a cell at all is on the what earns a row page. Other findings: who can read the signal, reasons printed beside a mark.