Decisions

Anyone who wants to say how far a machine took part in a text finds a crowded field. Some twenty approaches stand side by side, several of them older than Maschinenschrift, most of them built for a field other than literature and well built for it. This page orders them by the question each one answers, gives the reason for every omission, takes one book through all the forms, and tells which of them Maschinenschrift tried itself and gave up.

1. 1. The problem
   Whoever says openly how much came from the model risks more than whoever stays silent.
2. 2. The answers in the field
   Some twenty approaches answer eight questions. Maschinenschrift answers a ninth.
3. 3. The omissions
   Thirteen times, what is right elsewhere is left out here, each time for a reason.
4. 4. The test case
   The same book taken through every form: most stay silent, flatten it or tell half.
5. 5. What was built and set aside
   The obvious was built more than once and given up again, and every decision against it is dated.
6. 6. The objection
   A notation normalises what it notates. The alternative is not less mediation but invisible mediation.

The problem

Many authors write with language models: for research, for polish, for drafts, sometimes for most of the sentences. Most of them do not say so. An open account can cost a contract, a prize, a reputation, and whoever stays silent usually gets through. Suspicion takes the place of the missing account, and it falls hardest on those whose prose does not sit in the middle of the curve.

What is missing is a form in which the author states herself how much of the sentences and how much of the thought in her book came from a language model, without the answer giving her away. Many have looked for such a form. The answers differ above all in the question they answer.

The answers in the field

The register orders the approaches by the question each one answers, and describes each one by what its own publications say. As of 28 October 2026. It is maintained; additions and corrections go to contact [at] maschinenschrift [dot] com.

Maschinenschrift answers a ninth question: what the author says, under her own name, about how much of the sentences and how much of the ideas in this book came from a language model.

The nearest neighbours are the labels and the process taxonomies. With the labels, Maschinenschrift shares place and address: a self-declaration on the work, addressed to readers. With the taxonomies, it shares the gradation and the responsibility that stays with the person. From both it leaves something out; what, and why, is set out under The omissions.

Outside this comparison lies the question of whether a text may be used to train models. It has its own signals and registers. Maschinenschrift does not address it.

Rachel So gives an overview of the whole field in A Survey on Artificial Intelligence Attribution Badges.

The omissions

Every form achieves something, and each one is missing from Maschinenschrift for a reason. The letters refer to the table above.

1. No detector (A). A detector answers quickly and without the author’s involvement; with mass-produced text that is a useful hint. But it reads only the surface and says nothing about where ideas and structure came from. Where it counts as proof, the burden of proof reverses: the author has to show that she did not do something. In Maschinenschrift the account comes from the one who knows.

2. No recording of the writing (B). A recording supports a claim, and in examinations at school and university it is in the right place. Maschinenschrift adopted one as the genesis register and withdrew it. There is also a principle: an author’s workshop is private. A recording changes the question from what do you stand behind to what can you produce. An author who wants a time trail of her own manuscripts keeps one outside the standard.

3. No signature on a file (C). A signature secures that a statement is unchanged and who made it; for images, agency reports and machine processing it is the right tool. A book appears as ebook, paperback, new edition, translation and audiobook, and each version is a different file. The declaration belongs to the work. A signature also shows who made a statement, not whether it is true. And the reader never sees it: it sits in a registry, not on the title page.

4. No review office, no certification, no fee (B, D). Where a third party stands behind a mark with a contract, an interview or a check, deception meets a hurdle. A voluntary mark cannot keep anyone out; it can give the honest a distinguishing sign. Fees and procedures hit first those who have the most to declare: self-publishing authors. And whoever checks holds power over the checked. The standard is built so that no one can fence it in, its founder included; where that rests on a self-commitment, the commitment is stated openly.

5. No yes or no (D). A label is a stance, declared openly and by name, and it has chosen the right procedure: assurance instead of measurement. Whoever holds that children’s books should come from human hands may print that on the cover. But the author whose outline a model smoothed, or whose sentences it helped to write, finds no place on a label; she is left with silence or rounding down. The scale contains the label: a book that rightly carries one stands at W 0 · G 0 and loses nothing.

6. No percentage threshold (D). A threshold such as at least nine tenths by people is simple and leaves some room. But it assumes a measurability that does not exist: a share of words, of ideas or of working hours is in each case something different. Maschinenschrift counts levels on two separate axes, and the author explains in her own words what the number means.

7. No tiered scale (E). Three tiers sort quickly, and for platforms and filters that is enough. But the cases do not lie on a line. A book with human sentences and a machine structure and one with a human structure and machine sentences land on the same rung. Two axes without an overall score say where the model was, and add up to no rank.

8. No task list (F). Where a reviewer needs to know whether the literature search or the data analysis was done by machine, a list of work steps is exactly right. The reader of a novel asks about the result: whose sentences she is reading, whose thoughts. Whether a prompt was short or detailed changes nothing in the book, and every additional field adds to the false precision.

9. No resolution down to the paragraph (F). For texts made of clearly separate parts it is precise. A novel is a whole; its sentences and thoughts are interwoven across all its chapters. The declaration belongs to the work; where parts differ, the narration says so.

10. No model name in the notation (C, F). Whoever compares tools needs the name. But models age faster than books, and a product name on the title page advertises in passing. Above all, the substrate is a medium, not an actor: the notation counts what came from it. An author who wants to name the tool does so in the narration.

11. No colour that judges (C, E). Colour makes a mark recognisable at a glance. But everyone reads green, yellow and red as a traffic light, and a traffic light passes judgement. A tiered scale from education replaced its traffic light with neutral colours for exactly this reason. The seal is black and white, a typographic notation like a hallmark: it attests a composition, not a value.

12. No compliance as purpose (C, G). Providers and platforms need legal certainty, and labelling duties and platform fields are built for that. Maschinenschrift documents and takes no legal position; using it fulfils no obligation. An account that goes only to the platform also never reaches the reader.

13. No co-authorship for the model (H). A model on the cover is the furthest openness there is, and an artistic position. But a byline is an address: it says who answers for the work. Maschinenschrift exists because the byline is the wrong instrument for this account.

Maschinenschrift shares more with its neighbours than it leaves out: self-declaration instead of measurement, a public registry, responsibility with the person, gradation instead of yes or no, open use without licence negotiation, and no detector as the foundation. What remains is a combination, not a unique feature: a self-declaration on the work, visible on the book, for readers, on two separate axes for surface and structure, with its own line for fact-checking, with a zero point that includes the labels, without review office, fee or infrastructure, built for literature.

The test case

La Scuola dei Silenzi is the first entry in the registry and the founder’s own book, published under the name I.O. Vargas. It is declared W 5 · G 3 · ausstehend: the sentences largely from the model; the originating idea, characters, rules and every decision from the author; the research structured through the model and not checked. The same book, taken through the forms of the field, as of 28 October 2026, each line following the definition the system itself publishes:

What was built and set aside

Maschinenschrift was itself made with a language model. Most of the sentences on this site came from the substrate, as they did in the novel the project began with. The history of the project therefore does not show that it was made without a model. It shows the decisions, which is what the Geist axis is about: which proposal stands and which falls.

The project built or adopted several of the forms in the field itself, and gave them up again.

In this field the obvious move is infrastructure: a registry, a review, a ladder with three rungs. Maschinenschrift built the obvious more than once. What sets it apart from its neighbours lies where it decided against it, and each of those decisions is dated.

This page declares itself, the only one of the core pages to do so: W 5 · G 3 · belegt. The sentences came predominantly from the substrate. A substantial part of the order came as proposals from there; which proposal stood, which name fell and what the page does not claim was decided by the founder. Every statement about a neighbour has been checked against the neighbour’s own publications, as of 28 October 2026. The seal stands at the end of the page.

The objection

Whoever provides a form makes the formed thing more visible. A clean notation for mediated literature normalises mediated literature, even with no one promoting it. From the side of the labels, which declare a stance, this objection weighs heavily, and it is not cleared away here.

The project accepts it, because the alternative is not the absence of mediation. The alternative is its invisibility. The comparison is not a mark against nothing, but an account against silence. The size of the undeclared mass has since been measured: The Flood and Its Bed.

How the seal is read is set out in the introduction; it can be put together in the seal generator, and the meaning of the axes is governed by the Foundation.