On a collapsed $2 million book deal
A debut crime novel lost its publisher this summer. The reported advance was around $2 million. The stated reason was not that the book was bad — nobody has said the book was bad. The reason was that the author’s agents said they could no longer authenticate how the manuscript had evolved. No revision history. No developmental notes. No false starts. The book existed; the trail behind it didn’t.
I am part of an organization that sells AI tools to authors. So you would expect me to be defensive about this, and I’m not. What I want to do instead is say plainly where we think the line is, because the industry is about to draw it in a place that punishes the wrong people — and because any company selling into this moment owes its clients a straight answer rather than a shrug.
The failure was verification, not writing. It matters enormously that the manuscript was unverified rather than rejected. Those are different findings. One is a judgment about a book. The other is a judgment about a filing system.
Apply the filing-system standard backwards and watch what happens to the shelf you already own. Writers burned drafts. Writers wrote clean. Writers dictated, worked in notebooks that went in the bin, drafted on machines they later sold, wrote in one language and translated themselves into another. None of them were collecting evidence, because nobody had yet invented the accusation.
The accusation now exists and is being applied retroactively — and the people least equipped to answer it are the ones without institutional scaffolding around them. The first-time author. The writer outside London and New York. The person who wrote the book on a phone during a commute instead of in a workshop that generates a paper trail as a by-product. An MFA cohort produces documentation automatically. A nurse in Wyoming writing at midnight does not.
That is not a quality filter. It is a class filter wearing a quality filter’s coat.
Post-hoc detection doesn’t work, and the fallback is worse. Detection tools cannot reliably identify a manuscript that was assisted and then genuinely revised. They also flag human writers whose sentence rhythms happen to run tidy and formal — which, in practice, means non-native English speakers and anyone trained in technical or academic registers get caught disproportionately.
So the industry falls back on process documentation. But documentation is trivially fabricated by anyone acting in bad faith. A person who actually intends to pass off machine output as their own can generate a convincing sequence of drafts in an afternoon. The person who can’t produce a trail is, almost by definition, the person who wasn’t planning to be audited.
The screen catches the naive and the honest. The deliberate walk straight through it.
What we actually sell, and what we don’t. Here is our position, stated as concretely as I can, because vagueness is how this industry has been avoiding the question.
We do not sell a tool that writes your novel. We have deliberately not built one, and it is not a gap in the roadmap.
What we build sits either side of the sentence, never on it. An assessment tool that scores a finished manuscript on hook, voice, pacing, structure, and publishing path — a second reader’s opinion, delivered in ten minutes instead of ten weeks. A diagnostic that tells a writer why the middle isn’t working. A conversion tool that takes a manuscript the author has already written and renders it in screenplay form, so a novelist can see whether their book has a film in it without paying for a spec adaptation first. None of these produce prose that a writer could plausibly present as their own composition, because none of them produce prose at all.
That distinction is doing real work. A structural note is not a ghostwriter. It never has been — an editor who tells an author the second act is inert and the sister should be cut has intervened more consequentially in a finished novel than any model suggesting a sharper verb, and we have never described those books as co-written. We describe them as edited, and the author’s name on the cover stays honest, because authorship has always been judgment and responsibility rather than keystroke provenance. The author took the note. The author chose which scenes survived. The author signs the thing and answers for it.
That definition doesn’t break when the note arrives from software. It breaks only if you have quietly swapped authorship-as-judgment for authorship-as-manual-labour — and manual labour was never the thing anyone was paying for.
The part where I concede something. The models these tools run on were trained on books, including novels, without their authors’ permission. That is not an allegation — it is established fact, litigated and partly paid for: Anthropic settled author claims over pirated training data for $1.5 billion, while a court held that training itself was fair use. I’m not going to pretend that away in a piece arguing for honesty. The unresolved question is not whether it happened but what a fair licensing regime looks like going forward — and that question belongs at the model layer.
But notice that this is a grievance against how the tools were made, not against a novelist who used one. Collapsing those two lets the companies that scraped the corpus off the hook entirely while the individual writer absorbs the punishment. The author of that crime novel did not train anything. Whatever the industry’s argument with the labs is, running it through a debut novelist’s career is the wrong venue.
What a defensible standard looks like:
Disclose up front. Put the tool question in the contract, at signature, the way an academic declares funding — with consequences for lying, rather than consequences for failing to prove a negative two years later. We’re adding this to our own editorial agreements and we’ll publish the language.
Pay the source. Push the licensing fight to the layer where the taking happened.
Then judge the book. Publishing has always had a mechanism for rejecting weak work. It’s called reading it.
We work with authors who have spent four years on a manuscript and cannot afford a $6,000 developmental edit to find out whether the four years worked. Giving them a fast, cheap, honest read is not a threat to literature. Telling them their book is unpublishable because they didn’t keep their drafts might be.
