The Door Out

More creative than the average

In January 2026, Scientific Reports publishes the largest comparative study to date of human and machine creativity. A group around the neuroscientist Karim Jerbi (Université de Montréal, Mila) tests nine language models — among them GPT-4, Claude, and Gemini — against 100,000 people: association tasks, plus short writing forms such as haiku, synopses, flash fiction. The finding has two halves. The models surpass average human performance. And they stay below the more creative half of the participants — the study speaks of a ceiling, “a ceiling that current LLMs still fail to surpass.” The press echo kept mostly the first half: “AI Is Now More Creative Than the Average Human.” Both halves are right, and they explain each other — not through another measurement, but through a geometry.

Surprise is relative

How much surprise a text contains is not a property of the text. It is a relation: surprising for whom. Every reader measures what she reads against what she herself could have expected — against her own model of prediction. The stock phrase contains nothing because it is expectable for everyone.

The Sprachsubstrat — the language substrate: what a language model supplies in statistically probable formulation — is the condensation of billions of texts: the competence of the collective, made retrievable. Its output is expectable for that collective and surprising for the individual whose model is narrower. Surprise is the distance between the collective competence and one’s own.

Hence the contradictory reactions to the same text: astonishing to one reader, flat to another. The much-lamented “AI aesthetic” is not a text grown worse; it is a reader who has outgrown the collective mean. And hence the first half of the finding. A procedure that retrieves the median of everything ever written surpasses the median of individuals — no one is measured against a machine; each one is measured against all.

Decompression, not creation

The second astonishment: that so much coherence falls out of a short prompt. It, too, deceives about the source. The prompt does not generate the information; it retrieves it — a key, not an origin. A short address fetches a whole article; the article was never inside the address. The substrate makes a space of possibilities walkable that had long lain in the language: distributed across everything ever written, located in nothing. Concepts become places, relations become directions; the analogy that astonishes always lay in usage — it had simply never been spoken. The surprise is decompression, not creation. This space is no eternal realm of forms. It is the fossil of past usage.

Two novelties

The texts actually written are a thin sample of this space. The substrate interpolates all of it; most of it no human has ever visited. An output can therefore be new — never before instantiated — and at the same time not new: already implied. Combinatorial novelty is a new point in the old space. Generative novelty is a new dimension: a relation, a stake the corpus does not contain.

For the second, an outside is needed. A new fact from the world, or a new question put to it — neither lies in the space; both come in through the person at the desk. The substrate maps the interior of the space without gaps. The human is the only door out.

With this, the second half of the finding reads itself. The most creative participants are not ahead because they combine faster — the substrate combines faster. They are ahead because they bring what the space does not hold: their own question, the fact from outside, the stake. The ceiling the study speaks of is the edge of the space.

Surprise plus recognition

What remains is quality — that good outputs do not merely surprise but land. Valuable surprise is not mere unpredictability; that would be noise. It is surprise plus recognition: one did not see it coming, and the moment one sees it, it is right. That is the structure of metaphor, of analogy, of insight. The substrate hits it when its deviation falls on a valid connection never drawn — the connection lay, really, in the geometry of usage. Chance also surprises; it is not right. The difference between a good output and chatter is whether the deviation hits something that holds.

The door out

“More creative than the average” is therefore neither an alarm nor an all-clear, but a statement of place. The substrate fills the interior of the space more densely than any individual ever could — and stays inside it. Whatever a work contains beyond that came through a door only a human opens. No one can tell from the finished text which door it was — only the one who sat at the desk knows. From here, then, no path leads to the detector, but one does lead to the account: the seal declares how much of the linguistic surface and how much of the ideas come from the space — and the reader knows where to look for the truly new.

More on the form: Idea. Why the project says Sprachsubstrat rather than machine or LLM: Substrate, not Machine. Why it refuses the detector: The Witch Trial.

Self-declaration of this note: W 5 · G 3. Core ideas from a workshop sketch of May 2026, distilled in dialogue with the substrate; the anchor mapping and most of the linguistic surface substrate-generated; selection, review, and responsibility with the editor.