Which scope claims a reader can actually apply
Four entries state how far their marking reaches. Two bound the claim by something a reader can check against their own files; two bound it by something only the vendor can see. As of 2026-09-12.
| Point | Where it sits | What the record holds |
|---|---|---|
| Bounded by a named feature | One entry | All images created with the stated editing features |
| Bounded by which model ran | One entry | Models from elsewhere are generally not marked |
| Bounded by a product family or a version | Two entries | No product names; a version number with no date |
Inclusion rule. Entries stating how far their marking reaches. Entries making an unbounded claim, or none, are not listed here. Order. Claims a reader can apply first, then claims they cannot.
1Applicability is the property that matters, not breadth
Somebody holding a folder of assets can ask which came out of a named feature and sort them. Nobody can ask which of their files came from a vendor's generative consumer products as a class.
That is not a failure of the wider claim. A layer applied beneath several products is the right shape for a duty addressed to whoever builds systems, and the wrong shape for sorting a delivery folder.
2Two of the four narrowings are invisible from the headline
A vendor hosting other people's models narrows by model. A vendor attributing credentials to a version narrows by date. Neither narrowing is legible from the sentence a reader is most likely to quote.
Both are hedges the vendor chose to publish, and both are the most valuable sentences on their pages. A register that tidied them away would report two products as marked and leave a reader unable to sort anything.
3What a reader should establish instead
Every scope question in this record resolves to one of two facts about a specific file: which model produced it, and which route it left by. Both are answerable at the time and nearly unanswerable later.
Which is the same conclusion the rest of this record keeps arriving at from different directions. Two fields written down at delivery answer more than any amount of reading afterwards.
- Microsoft Designer (microsoft.com)Content credentials, provenance based on the C2PA standard, are applied to all images created with the editing featuresan image generation and editing tool
- SynthIDAn invisible watermark is embedded across Google's generative AI consumer products, added at the moment content is createdthe marking layer behind Google's video models
- ElevenLabsModels from other providers offered through the platform are generally not watermarked unless stated otherwisecoverage inside the vendor's own product
- Amazon Nova ReelVersion 1.1 adds Content Credentials based on the C2PA specification, carrying the model and the platform usedthe second provenance layer
4Sources
Article wording from AI Act Explorer, Article 50; vendor statements from the pages listed on each generator entry. Both read 2026-09-12. Related: What an upload leaves, Naming a standard, What buying changes.