Maschinenschrift Open standard for the self-declaration of language-substrate-mediated literature.

Introduction

Three weeks in May 2026

The Commonwealth Short Story Prize is one of the major English-language awards for short fiction. In 2026 it received 7,806 entries; a jury of five chose five regional winners. The Caribbean award went to The Serpent in the Grove by Jamir Nazir of Trinidad and Tobago. The story was published by Granta.

Days later, a business professor ran the text through Pangram, a detector for machine-written language. The result: 100 percent. The company then checked the winning stories of earlier years and published the numbers: in 2025, one of five regional winners was flagged; in 2026, three of five.

None of this proves anything. Detectors compute probabilities, not facts, and none of the accused writers has admitted to anything. What is remarkable is how the institutions responded. The Commonwealth Foundation stated that it does not check at all — the shortlisted writers had confirmed they used no AI, and it intended to honor that trust. Granta stood by the publication and had the text assessed by a language model; the verdict was that it was almost certainly not produced without a human hand. An institution for which AI is the scandal asked AI whether AI was involved.

That same month, the Nobel laureate Olga Tokarczuk told a conference in Poznań that she had bought herself the most advanced version of a language model and was often deeply shocked at how fantastically it broadened her thinking; for her new novel she had asked it what songs her characters might have danced to decades ago. A storm followed. A Polish colleague publicly likened the practice to marrying a vibrator; others called for her Nobel Prize to be withdrawn. Forty-eight hours later came a statement: she used artificial intelligence as most people do, as a tool for checking facts — “as a tool for faster preliminary research.” None of her texts had been written with its help.

And in May 2026 a detector classified parts of a papal encyclical on artificial intelligence as machine-mediated. The fact-checkers at Snopes reached a verdict: unproven.

Not one of these cases was settled. Each was decided — by suspicion.

Why detection does not help

Detectors do not answer the question being asked. What is asked is what a person did. What is measured is how probable a text is. Whoever writes unusually stands out — and unusual here means: far from the mean of the training material. This is why false positives fall with striking regularity on the same groups: people writing in a second language, authors from the edges of the English-speaking world, prose that was smoothed out at school. The Nigerian writer Innocent Chizaram Ilo, himself a Commonwealth prize winner, named the condition in one sentence: Everyone is an AI Cop Now.

There is also a technical problem that cannot be solved. Anyone who rewrites a machine-written chapter by hand defeats any detector within an hour. The race between detection and rewriting is structurally lost — for detection. What remains is a technique that waves the unremarkable through and punishes the conspicuous.

And even if detection worked reliably, its finding would be worthless on its own. The quick brown fox jumps over the lazy dog is a faultless sentence: grammatically correct, semantically clear, entirely indifferent. Whether a human or a program produced it changes nothing about that. Nobody reads in order to establish provenance. We read because someone meant something — because this sentence stands in this place, because a decision lies behind it, because someone answers for the claim. That is the difference between a good text and an arbitrary one, and it is not a property of the words but of the intent with which they were set.

Detection therefore does not merely measure the wrong thing; it measures the immaterial one. It establishes where a formulation came from. What matters is whether someone means it and stands by it. This is why the seal declares — not because machine provenance of words would be an offense, but so that the reader knows what relation she is in: what was chosen, what was supplied, and who answers for the whole.

The countermovement repeats the same error from the other side. Where publishers and platforms now offer labels certifying a work as guaranteed human, being human becomes a property that has to be tested and attested. Such a label is only as good as the procedure behind it — and there is no procedure.

The choice that is none

Anyone seriously intending to control the involvement of AI models in writing has two options.

The first is surveillance: log the keystrokes, record the screen, put a camera in the study. Only that would distinguish reliably. Nobody wants it, no legal order in Europe permits it, no publisher could pay for it, and after a weekend it would be circumvented — one simply copies out the machine text with random pauses.

The second is trust. It is the option every institution chooses, because the first is not one. But trust that can be lied to without consequence does not hold. Whoever answers honestly risks a contract, a prize, a reputation; whoever keeps quiet gets through. Under these conditions silence wins and the honest pay. The Tokarczuk case is the proof in its purest form: a Nobel laureate with the largest conceivable reserve of standing needed two days to withdraw a true statement.

That is the situation. What is missing is not technology and not morality. What is missing is a form in which the truth can be said without costing a life’s work.

What Maschinenschrift is

Maschinenschrift (German: machine-script) is an open standard and a mark. The author states for herself, in her own hand, what a language model contributed to her work — on three axes, in a notation that fits on a book cover.

  • Wortlaut (W) — wording; how much of the linguistic surface came from the model: sentence construction, word choice, rhythm. Scale 0 to 5.
  • Geist (G) — German for spirit, intellectual presence: how much of the ideas, arguments, and structural decisions came from the model. Scale 0 to 5.
  • Beleg (B) — the record; whether the work is bibliographically documented and findable. Two states: documented or pending.

The mark itself is two square brackets with three rows of five slots each. Dots count what the model contributed. An empty seal is the normal case: the author was alone. That direction is the most important decision in the whole construction. What is counted is not the human’s achievement but the machine’s share. Whoever writes without a model has nothing to demonstrate — she is the unmarked case.

The seal goes where the author’s name goes: title page, imprint, back cover. It is voluntary, costs nothing, requires no registration, and is approved by no one.

For comparison, here is what exists today instead. Anyone self-publishing on Amazon ticks a box stating whether the book is AI-generated or AI-assisted. Two boxes for the entire space between a model that looks up facts and a model that supplies the sentences. What does not fit in that box is the sentence that would be true of many books: The sentences are largely the machine’s; the characters, the structure, and the argument are mine. That sentence has a form in the seal. It reads W 4 · G 2.

What it is not

A text about AI and literature is sorted into a camp immediately. So, explicitly:

  • Not a purity seal. The mark does not certify that no machine was involved. It states to what degree one was.
  • Not a judgment of quality. It says nothing about whether a book is good. A seal is not a recommendation.
  • Not detection. Nothing is scanned, measured, or classified. Maschinenschrift does not detect. It records.
  • Not an accusation against those who keep quiet. A book without a seal is not a suspect book. The standard has no burden of proof to assign.
  • Not a verdict on AI. Maschinenschrift does not answer what models may be trained on, what the concentration of power in a few companies does, what running them costs in energy, or what the habit of delegation does to thinking. Those are serious questions. They are not this question.

This self-limitation is not caution; it is the condition under which the form stays usable for everyone. Whoever rejects AI models, whoever works with them, whoever holds both to be defensible — all three gain the same thing from mediation becoming documentable without the documentation turning into a test of conviction.

How it came about

The standard has a concrete origin, and an unspectacular one. In 2026 a novel was written with language models: the sentences came largely from the machine, the characters, the structure, and the material did not. When publication approached, the only available account was a checkbox on a retail platform. Nothing existed that could have expressed the actual state of affairs — no vocabulary, no scale, no convention. The choice was between silence, a tick mark, and a claim of purity.

The first version of the answer was an authority. A register with a submission form, a confirmation email, editorial approval, sequential numbers, encryption, annual review. It was built and it worked. In May 2026 it was scrapped. Two reasons: a standard that makes people ask permission does not get adopted — every day of waiting is a day someone drops it. And a machine that is a medium rather than an author needs no supervisory body; it needs a language. What remained is what had to remain: a notation, a specification, a registry, an email address.

The name is older than the project. In the twentieth century, Maschinenschrift meant the product of the typewriter — text that had passed through a technical apparatus, written by a human, mediated by a machine. No one ever claimed the typewriter was the author. The same distinction is now needed one stage further on.

So that none of this can later belong to one person, it was given away before it had value: content and notation under Creative Commons, the code under a free license, the word mark left dormant. Whoever uses the standard today cannot be cut off from it tomorrow — not even by its originator. The registry lives in text files, not in a database somebody has to operate. There are no accounts, no cookies, no trackers. Running it costs about as much per year as one dinner.

The novel it all started with declares W 4 · G 2 · B pending — much machine in the language, little in the Geist, no bibliographic record, because fiction documents nothing. That is the most uncomfortable account this project had to give, and it stands at the top of it. Whoever proposes a form for honesty has to answer first.

The objection that remains

One can lie in a seal. That is true and cannot be prevented.

The seal is not a proof but a statement — closer to a sworn declaration than to a notarized certificate. Whoever gives it stakes their reputation on it. Whoever lies in it lies, and can later be shown to have lied, because a checkable claim stands in the world instead of a diffuse silence. No procedure does more, and a detector does less.

The second objection is finer: whoever declares because declaring earns trust has calculated. Honesty as strategy cannot be told apart from honesty — not from the outside, at least. That too is true, and it is the reason the account is declared rather than tested: because no procedure separates the two cases, what remains is the statement of the one who knows.

What Maschinenschrift offers is therefore narrow, and it fits in one sentence. Not the question did she use a machine — but the question what does the work itself say about how it came to be. Where that question has a form of answer, suspicion has nothing left to expose.

The standard demands nothing. It provides a form.

The mark for your own work can be assembled here: Seal. The entered works are in the Registry, and admission runs by email. For the terms in more detail — why the project speaks of mediation rather than collaboration — the long version is under Idea; the running responses to incidents and texts are in the Notes.

Sources