The Flood and Its Bed

Two decades, one summer

For two decades Amazon passed for the book market’s problem. Since this summer, Amazon’s book market has one of its own.

In the summer of 2026 a research team around Tuhin Chakrabarty — Stony Brook, Columbia, Michigan, MIT — delivered the first full survey of what language models have done to self-publishing. The title of the study, as yet unrefereed, means its image literally: Generative AI floods and dilutes the market for books. The data: 14,419 self-published genre e-books on Amazon, January 2023 through March 2026, with sales tracked into June.

The finding

In three years the catalog grew thirty-eight-fold. Revenue grew nine-fold. The difference is the dilution.

Books the study classifies as predominantly machine-generated make up a fifth of the catalog. Most of them barely sell; a narrow top, however, together with the lightly mediated titles, now holds roughly two fifths of the highest sales ranks. At the same time revenue per title is falling even for books with no machine mediation at all, in six genres out of eight. What displaces is volume, not quality — the study says so itself: without competing on quality at all.

With that, the study documents the diagnosis this project stands on. And it ends at exactly the place where the project begins.

The bed

A flood needs a bed, and this one was dug long before the water came. Kindle Direct Publishing took the publisher and the up-front costs out of publishing. Kindle Unlimited pays not for the book but for the page read. Both reward the same thing: many titles, quickly replenished. What was missing was a substance that could be poured without limit. The Sprachsubstrat — the language substrate, the material a language model supplies as statistically probable phrasing — delivered it. The production cost of a novel falls toward zero, and the bed fills. The study shows the fit in a single number: in the genres that lean hardest on the page subscription, unmediated books lose the most ground — eight and a half points of sales share.

By now the bed is being built on purpose. A platform like NovelMint has serial fiction generated chapter by chapter, five cents per chapter read, some fifty dollars of production per book; the readers’ reactions feed into the next run. A label for the mediation is not provided for. Where everything is mediated, it would distinguish nothing. The flood there is no longer an inflow muddying a body of water. It is the water.

Wording is cheap

The study’s most instructive finding is a linguistic one. Successful mediated titles overlap in rare turns of phrase markedly more with existing works than unmediated ones do — by seven percentage points in the best-selling tier — and the greater the overlap, the greater their revenue. For unmediated books no such relation holds. Prize-winning literature overlaps least of all.

The seal of this project counts what the machine supplied, on three axes: Wortlaut — wording, how much of the linguistic surface came from the substrate. Geist — how much of the ideas, the arguments, the structural decisions, the factual material, the content. Beleg — the record, whether the work stands in the bibliographic world. The study gives that order an empirical floor. Wording is the axis on which the substrate delivers most — and what it delivers is borrowed: statistical proximity to what was already written. What lifts a work out of the flood sits on the other axes — in the Geist that was not imported from the machine, and in the record that fixes the work in the bibliographic world.

The unmeasured quantity

The study notes itself what its data cannot yield: whether readers choose differently when they know about the mediation. It could not be measured, because in this market there is nothing to know. Amazon has asked every submission since 2023 whether it contains machine-generated content; the answer stays in the house and appears on no product page. The account whose effect the study must leave open exists nowhere in this market in public.

The flood is therefore to be named more precisely. It does not consist of mediated books; it consists of mediated books that do not say so. The flood is the undisclosed mass. Self-declaration is the counter-movement to this state of affairs: public instead of in-house, graded instead of ticked, given by the author instead of asserted by a tool. Whether the open account changes how readers choose is one of the open questions of this decade. The study shows that so far no one has had the chance to answer it.

In the aggregate, at the single work

The study measured with a substrate detector — a tool that scans texts for machine mediation. For surveying a market, that holds: in an aggregate of fourteen thousand books no one stands accused, and the errors cancel in the statistics. At the single work the same tool would remain what it has always been: a witch trial that tests provenance and passes over sense. Whoever wanted to build a gate out of this study — a detector at the market’s entrance — would get the familiar selection: the guileless caught, and whoever washes her traces waved through. A market can be scanned. A book wants to be read.

Signal, not gate

A registry stops no flood. It never promised to: signal, not gate. What a registry can do is record how a work was made and who stands behind it. The flood has no sender; that is what makes it a flood. Whoever declares puts a name where otherwise there is only volume — no acquittal and no exit ticket, but an account given to the reader. What she makes of it is hers to decide.

And the registry infers nothing from the absence of a declaration. It does not read silence, does not comment on it, and holds it against no one.

Amazon dug the bed, the substrate filled it, and the next beds are under construction. The water level is not this project’s concern. The account is.

On the form and the line: Idea. Why the single work is declared, not scanned: The Witch Trial. The derivation of the term: Substrate, not Machine. The more precise version of the sentence that wording is cheap: What Cannot Be Accelerated.

Self-declaration of this note: W 5 · G 3. Occasion, hook, and position with the editor — the study from the editor’s reading, the reversal of the flood’s bed from the editor’s call. Linguistic surface substrate-mediated; figures checked against the study; selection, review, and responsibility with the editor.