← Build Log

What We're Actually Arguing About When We Argue About AI Writing

April 17, 2026 · Revised September 15, 2026

The week I paid attention to Twilight

I was sick for a week, stuck on the couch, and ended up watching Twilight — not my normal fare, but it seemed harmless enough, and I did briefly wonder whether the cold medicine was a factor.

I noticed something. Almost nothing happens. Two people meet. They stare at each other a lot. There's a diner, a field, a hospital, a couple of houses. The plot fits on an index card. And yet the central characters drew me in. I was interested.

When it finished I wanted to know why it had worked, so I tried The Vampire Diaries, which is supposed to be the same thing. I turned it off in under thirty minutes. The vampires clearly weren't the deciding factor.

Something specific had worked in the first case and was absent in the second, and it wasn't the concept. It was the craft. That's the kind of question I'm drawn to, and the rest of this project grew out of that week. It's also, as it turns out, the question this essay is about.

The question this essay is about

I'm not going to argue that AI is good or bad for writing. That's the wrong question and it produces the wrong conversation. The narrower, less comfortable question is: when we treat AI writing as the bogeyman, what are we actually afraid of — and is the fear pointed at the right thing?

I think it isn't. I think the industry is fighting a battle over whether a machine touched the manuscript, and losing the battle that matters, which is whether we can still evaluate a book on its craft.

The case that changed the conversation

Earlier this year a major traditional publisher cancelled a contracted debut novel over allegations of AI involvement. I won't recap the details — the dispute will be settled wherever such things get settled, and the author has contested the claims. Whatever the outcome, the signal to every agent, editor, and debut author was unambiguous: the appearance of AI involvement is now a career-ending liability, and the safest institutional response is zero tolerance.

Now consider the industry's own answer to that climate. The Authors Guild launched a "Human Authored" certification in January 2025 and opened it to all US authors in March 2026. Authors license a trademarked mark declaring their book was written by a person. The Guild has been candid about how it works: the certification rests on the author's signed attestation, backed by a licence agreement and the threat of liability — not on any verification of the manuscript. In the Guild's own words: "As of now, we are not aware of any truly reliable way to test whether a work includes AI-generated material." Its CEO put it more directly: no AI detection tool exists today that is sufficiently accurate for the purpose.

Credit the Guild for saying so plainly; most of the industry hasn't. But sit with what it means. The organisation representing working authors, launching a programme to certify books as human-written, states that it has no way to check.

That admission is the whole problem in one sentence. If no reliable method exists for telling AI prose from human prose, then every decision made on the basis of suspected AI involvement — every cancellation, every quiet pass on a manuscript, every whisper network — is being made without an evidentiary standard. And the seal offered to reassure readers means exactly what an author's signature on a form means.

Neither the suspicion nor the seal tells us anything about the craft of the books involved. Neither distinguishes between a novel generated by a language model and a novel where a writer used AI as a drafting and revision instrument. The conversation has collapsed that distinction without acknowledging it exists.

What ghostwriters tell us about authorship

Consider a set of books nobody argues about.

James Patterson publishes dozens of titles a year, most drafted by co-authors whose names appear on the cover in smaller type. Nobody calls this fraud. Most celebrity memoirs and political autobiographies are ghostwritten; the named author may not have written a sentence of the manuscript. Stephen King published as Richard Bachman for years, and the Bachman books didn't become less valid as literature when the pseudonym came out. The industry has a large, stable set of precedents for authorship when someone other than the named author did much of the writing.

The criteria are roughly these. The named author made the creative decisions: what the book is about, what the voice sounds like, who the characters are. They read and approved every line. They are the person who lives with the finished work professionally, legally, and reputationally. The person who did the drafting worked within that creative authority. They were an instrument of it.

By those criteria, an AI-assisted book where the writer designed the world, built the characters, locked the series architecture, approved every line, and rewrote extensively is authored by that writer. The AI is the instrument. The writer is the authority. This is the structure ghostwritten books already operate under, unquestioned, across the industry.

If the objection to AI-assisted writing were really about authorship, the ghostwriter comparison would be fatal. Which means the objection is about something else.

Where the analogy breaks

The parallel isn't perfect, and an honest argument names its own weak points.

A ghostwriter is a person, compensated for their labour, who chose to take the work. Language models were trained on material that included copyrighted work whose authors were not consulted or paid. That's a real grievance, at the centre of active litigation, and any honest account of AI writing should say plainly that the economic model these tools rest on is still contested.

There is also the labour concern. If these tools let fewer writers produce more books, working writers lose income. That's true, and it's painful for the people it lands on. It also belongs to a recognisable historical category — the same thing has happened with every productivity technology in every creative field — which doesn't make it less real, but does make it a question of policy rather than of fraud.

And there is the cultural anxiety. A lot of what people mean by "AI writing is bad" is "I'm afraid of what this does to the idea of humans as creative beings." That's a real feeling and it deserves to be taken seriously. But it isn't the same as "this book is fraudulent because a machine helped draft it," and collapsing the two produces worse outcomes for everyone, including the people voicing the concern.

The legitimate objections are training data and labour. The objection used in public — this book is fake because a machine helped — is neither. It's a proxy for the real concerns, dressed up as a craft argument, and because it's dressed up we can't address it directly. We can only argue about books.

Zero-tolerance policies on AI involvement aren't about training data, and they aren't about labour. They're about appearance. And any decision made on appearance requires the publisher, and the public, to be able to tell AI-generated prose from prose that isn't — which, by the industry's own admission, nobody can. That isn't an editorial standard. It's risk management, and it's worth being honest that that's what it is.

What AI actually did to my project

I started the Nascent series because I wanted to understand what was possible. I built characters, designed a world, roughed out a four-book arc, and used AI for idea generation and early discussion — which is what most people mean when they say they use AI for writing.

That would have been the whole story, except I work in a quantitative field where AI workflows are standard, and it was obvious this was the floor, not the ceiling. So I built architecture: character documents, a world-building bible, a series arc, tone guardrails, a chapter-by-chapter beat sheet, an epigraph progression, a chapter-title grammar. Dozens of documents and tens of thousands of words of scaffolding before a chapter of manuscript was drafted.

With that in place, I could interact with the entire four-book story in real time. Ask about character interactions I hadn't written. Run hypothetical changes and see which beats they broke. Seed an idea at chapter three and trace its implications through chapter twelve without having written either. Hold the whole shape of the series in working memory and examine it from any angle.

This is the part I want people to understand, because it's the part the current conversation is missing.

The master storyteller's research layer

Novelists who write sprawling, internally consistent series don't manage that consistency by accident. They build infrastructure: character bibles, world-building notebooks, maps, timelines, relationship charts. Terry Brooks has built Shannara across four decades. Brandon Sanderson runs what amounts to a small company of assistants, beta readers, and continuity checkers. The scaffolding that makes ambitious fiction possible takes years to build and is invisible to readers.

AI gave me real-time access to that scaffolding on day one.

Yes, it also gave me drafts — a full first draft, in fact, and I've written elsewhere about what those drafts were worth and how many passes it took to make them mine. Drafts are the cheap part. The model can now produce competent, physical, specific prose when it's given enough of the writer's own established material to draw on. What it can't do is select — pick the one detail out of twenty that's alive, the one only this narrator in this moment would notice. That remains the writer's job, and it's the job that makes the book worth reading.

So the debate's fixation on whether the machine can generate a good sentence is aimed at the wrong thing. What the machine does, given the right architectural input, is make the research, continuity, and consistency layer of serious fiction navigable for a debut novelist in a way that used to require years of back-office work and possibly a staff.

The prose in my case is substantially rewritten. Every line passes through my hands multiple times before it stays. The voice, the architecture, and the decisions are mine. What AI contributed was the ability to hold the whole project at once — which is what novelists with twenty-year careers have done the long way around since the novel was invented.

If that's fraud, the category is much larger than anyone is willing to admit.

What we should be talking about

The standards for evaluating a book should be the standards for evaluating any book. Is it good? Is the voice consistent? Are the characters earned? Does the story work on its own terms? Is the author accountable for the finished product?

Those are the questions that matter, and against them the question of which tool produced first-draft prose is a process question — in the same category as which word processor the writer used, or whether they dictated, or whether a developmental editor reorganised three chapters.

That does not make disclosure irrelevant. It's the ghostwriter precedent again: Patterson's co-authors are on the cover. Whether AI was used says nothing about whether the book is good, but readers are entitled to know how the thing they're holding was made, and a professional culture that punishes honesty gets less of it. Which is what we have. A 2025 BookBub survey of more than 1,200 authors found that about 45% were using generative AI somewhere in their work — and of those, 74% do not disclose it to readers. Call that what you like. I'll call it a culture where everyone has agreed to pretend together.

I'd rather be open about it. I should also be honest that openness costs me less than it would cost a traditionally published debut author with an agent and a contract to protect: I'm self-publishing, under a pen name, and no one can cancel a deal I don't have. That's exactly why the current climate is so corrosive — the people who can afford candour are the ones who need it least.

Final thoughts

I'm not asking anyone to approve of how I'm writing this series. It's been one of the most enjoyable things I've done at the intersection of technology and craft, and I'm happy to share it. I'm asking that if you disapprove, you notice which question you're answering.

If the question is is this craft? — then it has to be evaluated book by book, on the book's own terms.

If the question is is this fair to working writers? — that's legitimate, and it deserves a real conversation rather than a certificate.

If the question is does AI-assisted mean AI-generated? — no, and the refusal to distinguish the two is how a debut novel and a cancellation end up in the same sentence.

If the question is am I afraid of what this means about being human? — I understand that, and I don't dismiss it. AI will keep posing questions we can't hide from. But that is a separate question from the other three, and its answer won't be found in a publisher's policy department.

The question isn't whether AI belongs in fiction. It's whether we can tell the difference between the work of the tool and the work of the writer. If we can, the category holds. If we can't, we have a much bigger problem than AI.

Quinn Frost is a pen name. The books are self-published. The How I Work page on this site explains, in more detail than any disclosure requirement would ask, how the sausage gets made. From there the reader decides — not a controversy that isn't about the thing it thinks it's about.

Sources