Synthetic media law, tracked to the article

SenseDefend records what published instruments say about generated audio, image and video: the clause, the role it binds, the date it starts to apply, and the wording itself. Entries carry the source they were read from and the date they were last checked. As of 2026-09-12.

Inclusion rule. Generators whose own public pages say something about marking generated output. A generator that publishes nothing is not given a row. Order. Alphabetical by generator.

1Two terms defined against the text

Two words that recur across the instruments have entries of their own: content moderation, aigc label. Each is defined against published material, never against how a product is described in marketing.

2What is recorded here

Each entry is an instrument read at the level of its numbered paragraphs. A duty is recorded against the paragraph that imposes it and against the role that paragraph names, because those two things decide whether a given obligation is anyone's problem. Dates are recorded as the instrument states them, with the clause that sets them; where an instrument is silent, the entry says so rather than filling the gap with an inference.

Wording matters more than paraphrase in this area, so passages that carry the operative language are quoted rather than summarised, with a link to the page they were read from. Everything else on a page is written plainly, and every number or date carries its own check date so a reader can see how old it is without reading the whole page.

3Current entries

• EU AI Act Article 50 — transparency duties for systems that interact with people or generate synthetic content, paragraph by paragraph, with the application date and the exemptions.
• Duties by role — which paragraph binds the provider, which binds the deployer, and the moment each duty takes effect.
• What generators document about marking — sixteen AI tools read against Article 50(2), from their own public pages.
• SceneMixer, read against Article 50 — a visible mark in a compliance guide, and the machine-readable half left unaddressed.
• Google SynthID — the mirror image: invisible, at generation, with no removal documented.
• Kling AI — the vendor calls it a brand watermark, which is a different thing from a disclosure.
• LTX Studio — read from a plan feature list, which is weaker evidence than a sentence.
• Pika — the mark leaves with a purchase rather than with a switch.
• Runway — the free-tier position is an omission on a plan table, not a statement.
• OpenAI Sora — signed metadata on every asset, a moving visible mark on downloads, and detection kept in-house.
• Adobe Firefly — a credential attached to outputs, with nothing said about a mark in the frame.
• Amazon Nova Reel — two layers at once, and only one of them needs the vendor to read it.
• Microsoft Designer — credentials scoped to images made with the editing features, read on a separate site.
• Synthesia — two unrelated conditions put the same mark on a file, and one of them is irreversible.
• HeyGen — removal named on two paid tiers and absent from the free one.
• ElevenLabs — audio, which the paragraph names and the discussion usually forgets.
• Meta AudioSeal — a scheme published as research, locating a mark rather than merely finding one.
• Alibaba Wan — the wording of the mark is published, and the same family also ships as downloadable weights.
• MiniMax Hailuo — the lever is a request parameter that defaults to off.
• How entries are recorded — what earns an entry, where the wording is read from, and what the two dates on every fact row mean.

4Questions this record can answer from the text

5Explainers

Background pages on how marking and record-keeping work in practice. They assert nothing about any product or any instrument; the sourced wording is on the generator table and the instrument pages.

• Three kinds of marking — metadata, embedded signals and visible labels are three mechanisms under one word.
• Keeping a record — what to write down at delivery so a question a year later is answerable.

Each of these starts with the answer in a sentence, then shows the wording and the vendor statements it rests on. None of them concludes that anything does or does not comply.

• A visible mark and a machine-readable mark — different paragraphs, different parties.
• What it means when a mark can be removed — a switch, a plan, and what the article does not say.
• What free tiers publish about marking — the least-documented tier carries the most output.
• Every carve-out written into Article 50 — only one of the three switches a duty off.
• When Article 50 starts to apply — the date lives in a different article.
• The points no generator addresses — interoperability, accessibility, timing.

6Article 50 paragraph by paragraph

Each numbered paragraph has its own page, carrying the same four points: who it binds, what it asks for, the carve-out written into it, and when it starts to apply. A reader who needs one duty does not have to read the whole article to find out whether it is theirs.

• 50(1) — systems built to interact with people, and what counts as obvious.
• 50(2) — machine-readable marking, and the one adjective a vendor cannot deliver alone.
• 50(3) — emotion recognition and biometric categorisation, a deployer duty.
• 50(4) — deep fakes, and the narrowing for evidently creative work.
• 50(4), text — published text, and the editorial-responsibility carve-out.
• 50(5) — how and when every notice has to be given, including accessibility.
• 50(7) — codes of practice, where the technical detail is expected to arrive.

7How entries are kept

An entry changes when the instrument or the page it was read from changes. Check dates move on every review, whether or not a value changed; the page date moves only when the content actually changes, so an date that has not moved says the page itself did not change, not that nobody looked at it. Where a reading is taken from a secondary compilation rather than the official gazette, the entry names that compilation.

Entries are added when an instrument carries operative language about generated audio, image or video, and the language can be read in a published version. Bills that have not passed, drafts under negotiation and commentary about what a rule might come to mean are left out until there is a text to quote, because a record of what the law says is only useful if everything in it is something the law actually says.

SenseDefend is a record of published texts. It is not legal advice, and it does not assess whether any particular product or production complies with anything.

8How the record is split

Split by source: official text on one side, vendor documentation on the other, never merged into a single verdict. The data page keeps the same split.

• Data — The recorded clauses and vendor statements as a comma-separated file, each row carrying…
• EU — The transparency article of the AI Act, recorded one paragraph at a time…
• Duties — The transparency obligations sorted by the party they bind, and what the generators…
• Generators — A page per generator…
• Fields — The fixed rows behind the tables here…
• Questions — Questions about synthetic-media transparency rules, answered by quoting official text…
• Terms — Definitions for the vocabulary these pages rely on…
• Learn — Explainers on the two distinctions this record depends on…
• Statements — Each page here takes a single generator on a single column of this record and says what…
• Pairs — Two generators set against each other on a single column, chosen because their…