Amplifier, not Co-Author

Who drives the story

At NLP4DH 2026, the conference for natural language processing in the digital humanities, Halfdan Nordahl Fundal and Yuri Bizzoni take the measure of writing in turns. Eighty-seven stories, written by participants alternating with one of four language models: ten rounds, each contribution around 25 to 30 words. Three questions: whether the two sides differ in tone, who aligns with whom — and whose moves remain effective as the story goes on.

Three findings

First: the model writes brighter. Its contributions sit measurably above the human ones in tone, and the pull toward the positive holds even against negative human moves. The authors’ suspicion falls on training and fine-tuning, in which strongly negative material is removed.

Second: the alignment is mutual but unequal. Model contributions follow the tone of the preceding human contribution more than twice as strongly as the reverse, and semantically humans stay slightly closer to their own previous contribution than to the model’s. The paper calls this resistance emotional autonomy.

Third, the central finding: narrative influence is asymmetric. It is measured information-theoretically — how surprising a contribution is against everything that came before, and how much of that surprise lives on in the contribution that follows. Human contributions are both to a higher degree: more novel and more lasting. Contributions of equal surprise shape the continuation more strongly when they come from the human; the paper calls that writing privilege.

The authors’ reading is a division of labour: the human sets direction and the new; the models act as adaptive amplifiers that sustain coherence and elaborate what has been set. The discussion draws the sum: less as creative co-authors and more as adaptable stylistic amplifiers.

The demotion

The paper opens with collaborators in quotation marks and ends at the amplifier. Its vocabulary posits two writers; its measurement describes a medium. This demotion from partner to material is the move this project made in its own usage before anyone measured: it says Sprachsubstrat — the language substrate, the material a language model supplies as statistically probable phrasing — and means a ground that is worked on, not a counterpart that works along. What is new is the route. Here the demotion comes not from the theory of language but from distributions.

A second convergence lies beside it. The paper suspects the creative value of the procedure lies less in autonomous narrative development than in the ability to surface and select among divergent narrative branches. That, phrased as a measurement result, is the sentence this project’s order stands on: selection is a constitutive act of authorship. The derivations are independent — the paper does not know this project, and this project did not wait for the paper. Two routes, one place.

What a slope can carry

The project cites this confirmation. It does not build on it. The measurement sits at the level of the text: tone values, surprise, coupling. The declaration sits one level above: who does what with a text and what they stand behind. A position that a regression slope could confirm would be a position the next regression slope could refute. If the follow-up study, with longer contributions or a stronger substrate, finds the reverse asymmetry, nothing about the declaration changes. The seal asks who stands behind the work — and no corpus answers that.

Restraint follows in both directions. Whoever takes from the finding that the account is dispensable, because the human drives anyway, turns twenty moves of around 25 to 30 words into a verdict on books — and overlooks that precisely what was measured is what grounds the account: the constant shaping of tone and surface by the substrate. And whoever raises the observed division of labour into a norm mistakes the mean of a laboratory game for the limits of the spectrum. The seal knows the entry in which the substance comes predominantly from the substrate and the author curates. The paper gives no reason to hold that configuration inauthentic. It describes what happens on average, not what is permitted.

The amplifier’s characteristic curve

For the declaration, the most inconspicuous finding is the most concrete: the pull toward the positive. What the substrate returns is not neutral, even where it merely elaborates. The amplifier colours what it amplifies — measurably, persistently, against the human’s move. The seal of this project declares on two axes what came from the substrate: Wortlaut — wording, the linguistic surface. Geist — the ideas, the arguments, the structure. The valence finding is the most precise reason yet why the first axis remains worth an account where the second lies entirely with the author: even the work on the surface drags in a direction of its own. An amplifier is not a neutral element. It has a characteristic curve, and the curve is in the text.

The laboratory has taken the amplifier’s measure. The account that one was involved is given by no instrument. It is given, or it is missing.

On the form: Idea. What the substrate brings forth on its own — and where its space ends: The Door Out. The derivation of the term: Substrate, not Machine.

Self-declaration of this note: W 5 · G 3. Occasion from the editor’s reading; analysis and demarcation from the project’s workshop; linguistic surface substrate-mediated; findings and quotations checked against the paper; selection and responsibility with the editor.