A visible mark and a machine-readable mark
They answer different paragraphs. Machine-readable marking is what Article 50(2) asks providers for; a disclosure a viewer can perceive is what Article 50(4) asks deployers for. A product can do one and say nothing about the other. From 4 source pages, read 2026-09-12 to 2026-10-13; no maker pays to be listed.
| Point | Where it sits | What the record holds |
|---|---|---|
| Machine-readable, at generation | Article 50(2), provider | One entry describes a scheme |
| Visible on the deliverable | Article 50(4), deployer | Most entries describe one |
| Both | Two duties on two parties | Two entries describe both |
Inclusion rule. Kinds are taken from what the paragraphs ask for, not from product vocabulary. An entry counts in a row only where the vendor's own pages describe that kind of mark. Order. Provider duty first, then deployer, then both.
1The two are not stronger and weaker versions of one thing
It is tempting to rank them: an invisible signal sounds more serious than a logo in the corner, so a product with one might seem to have covered the other. The paragraphs do not read that way. They impose separate obligations on separate parties, and the party that has to disclose to a viewer is usually not the party that built the generator.
That has a direct consequence for anyone buying a tool. A vendor's marking scheme, however good, does not discharge the buyer's own duty when the buyer is the one putting content in front of people. Reading a supplier's documentation as if it settled both is the most common way to end up with a gap nobody owns.
2What the record actually contains on each side
On the machine-readable side, SynthID describes a signal embedded at generation and what it is meant to survive, and SceneMixer's terms describe an invisible watermark or content credential as compression-resistant. On the viewer side there are several entries, almost all of them plan features rather than policies.
Two entries describe both halves for the same product: Sora's system card, with a mark on downloads and C2PA metadata on every asset, and SceneMixer's terms, with a visible label and an invisible layer that is never removed, under rules that include Article 50 itself. Everyone else describes one half, which is the most useful observation a reader can take from this page, and it is an observation about published material rather than about what the products do.
3Why this page does not conclude anything
Whether any particular arrangement satisfies either paragraph depends on facts about a deployment that no register holds: what was generated, who published it, to whom, and in what context. This record keeps the wording and the vendor statements side by side so that the question can be asked precisely, and leaves the answer to the people who have those facts.
What it will do is say when the two sides are not comparable. A plan feature and a technical description are different kinds of evidence, and a table that puts them in adjacent columns can make them look alike.
- Article 50(2)Providers must mark synthetic audio, image, video and text in a machine-readable formatincluding general-purpose AI systems
- Article 50(4)Deployers must disclose image, audio or video content that is a deep fakefirst subparagraph
- SceneMixer (scenemixer.com)A small corner badge is applied at delivery, on by defaultan AI short-drama generator
- SceneMixerOutput carries an invisible, compression-resistant watermark or content credential under China's deep-synthesis provisions and EU AI Act Article 50, and the invisible layer is never removedmachine-readable marking
- 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
4Sources
Article wording from AI Act Explorer, Article 50; vendor statements from the pages listed on each generator entry. Both read 2026-10-13. Related: Marks that come off, What the free tier does, The carve-outs.