What the Law Does Not Require
A date and a setting
Since 2 August 2026, Article 50 of Regulation (EU) 2024/1689 — the AI Act — has applied. It governs the transparency obligations for certain AI systems: whoever speaks with a system should know it, and whoever puts out synthetic content should mark it as such. The European Commission published its interpretive guidelines days before the date of application. A breach costs up to fifteen million euros or three percent of worldwide annual turnover.
Some three weeks earlier, on 9 July 2026, Google amended its advertising policies. Ads that ran through Google’s own generative tools are labelled automatically; whoever uses other tools sets the label through a control of their own. In My Ad Center a panel appears alongside it: How this ad was made.
Two events, one subject. Between them it can be shown what a labelling duty achieves and where it stops.
What Article 50 requires
The article separates its addressees. The providers of the systems must mark their outputs in a machine-readable format — at the source, in the output itself, not in the finished text. The deployers must disclose what they put out when it is a deep fake: image, audio, video.
Text is not in that sentence. For text there is a sub-paragraph of its own, and a narrow one. What must be disclosed is text published for the purpose of informing the public on matters of public interest. A novel informs the public of nothing. It asserts nothing that holds, and it says so in advance.
The novel does not fall under Article 50. This is neither a loophole nor an oversight, but the correct line: the provision protects against deception, and where nothing is feigned, there is no one to mislead.
The exception for the invented
In one place the law says as much itself. Where content forms part of an evidently artistic, creative, satirical or fictional work, the duty shrinks to a disclosure made in an appropriate manner that does not hamper the display or enjoyment of the work. Not on the work, but beside it. Not in the text, but in the apparatus around it.
That is the remarkable place in the whole provision. The legislature wrote a labelling duty and conceded, in the same paragraph, that it damages a work of art when laid upon one. The Commission reads the exception narrowly — where informative and creative character are mixed, the informative prevails. The direction of the concession stands nonetheless. The warning label is built for the commodity, not for the work.
What Google does not do
Google commands the largest computing power ever turned on the inspection of advertisements. The company does not use it to detect synthetic advertising. It labels what has run through its own tools — there it knows — and for everything else supplies a setting the advertiser operates herself. The declaration is compulsory only where a jurisdiction demands it, and in election advertising, since 2023. The policy states expressly that the setting guarantees no compliance: the obligation stays with the one who places the ad.
The rest is an account. Whoever placed it says how it was made. This is not negligence but the only construction that holds — the machine detection of machine-mediated text has failed on its own terms, and whoever is answerable knows better than any detector anyway. The demonstration is set out elsewhere: The Witch Trial.
The largest advertising market on earth does not detect. It asks the sender.
Warning label and account
A label under Article 50 is a warning. It has an addressee — the consumer — a danger — deception — and a form: yes or no. It says: This face never spoke. This voice belongs to no one.
In literature the warning runs empty. Whoever opens a novel need not be protected from the news that something in it was invented. Fiction is the one kind of speech whose untruth has been declared in advance. The label AI-generated, which Amazon KDP has carried as a mandatory field at upload since 2023, imports the logic of the hazardous substance into a room where there is no hazard. And it leaves unanswered the only question a reader has about how a book came about: not whether, but how much of what.
The warning label asks whether a machine was involved. Self-declaration states in what proportion.
The seal is therefore not a checkbox. It carries three axes: Wortlaut — how much of the linguistic surface came from the Sprachsubstrat, the language substrate. Geist — how much of the ideas, the structure, the content. Beleg — whether the bibliographic record has been established. Two scales from zero to five and one binary state, where the law has a box. It does not accuse; it records.
Where the law is silent
The regulation is right where it sits. Commercial and political space need protection from language with no sender behind it; a deep fake in an election is an attack, and a warning label is the least that answers it. For the title page of a book it will not serve — and there the provision does not ask for it.
That leaves literature standing where it has always stood in the history of writing technology. The printer’s device sat on the title page before any rule required it. The colophon named place, year, and hand because these belonged to the matter, not because an authority asked. What is not owed is exactly what is worth saying.
The AI Act protects the consumer against deception. To the reader it owes nothing. What is said about how a novel came about is said by no one but the woman who wrote it — voluntarily, or it goes unsaid.
On the form in which the account is given: Idea. On the axes and their notation: Siegel-Spec v7.
Self-declaration of this note: W 5 · G 3. Conception, occasion, and position with the editor. The elaboration ran through two language substrates: the first supplied a version whose central factual claim — that Google compels every advertiser to declare on pain of account suspension — did not survive checking. The second checked it, discarded it, and rebuilt the argument. Legal position and linguistic surface substrate-mediated; selection, review, and responsibility with the editor.