Your Daughter Can’t Ride a Bike. Buy a Trailer.

A literary agent I read, Abigail Fenton, published a piece called “No, You Can’t Write Your Novel With AI.” It is well argued, it is honest about being personal opinion instead of industry policy, and I think it is wrong in ways that matter, but the part that stuck with me was the ending, because it gives away more than she meant it to.

She closes with her five-year-old, and the story goes like this: friends went on a family bike ride, ice cream at the end, and the daughter couldn’t come because she has never been willing to put in the time to learn to ride. “If you want to ride a bike, you’d better get in the saddle and learn.”

I have looked at that paragraph for a while now, because every parent I know has solved this exact problem. There is a child seat that bolts over the rear wheel, there is a trailer that hitches to the seatpost, there is a tag-along half-bike that lets the kid pedal without steering or balancing. Four options, roughly forty to three hundred dollars, all of them at your local bike shop, and with any of them the five-year-old gets the ride and the ice cream and, incidentally, five miles of watching how riding a bike actually works.

The analogy chose itself, and it does not say what she wants it to say. It says: when someone can’t do the thing yet, you can either build the thing that carries them, or you can let them stay home and call it a lesson.

The ghostwriter problem

Her strongest argument is legal, and it is the one I take seriously. AI-generated text is not copyrightable, so an author cannot grant a publisher exclusive rights to something they never owned, and publishing contracts contain a warranty that the work is your own creation. If a machine wrote it, you cannot sign that warranty. Fine, and that is an accurate description of where the law sits today.

But she then slides from a contract problem into a definition of authorship, and that is where it falls apart, because publishing already has a very large exception sitting in plain sight. Celebrity memoirs. Politician autobiographies. Business books by executives who have never written a paragraph in their lives. A professional ghostwriter does one hundred percent of the word-picking that Fenton says makes someone an author, and the byline still belongs to the person who supplied the life, the ideas and the voice. Nobody calls this fraud, because it is a category with its own rates and its own quiet etiquette, and it has coexisted with the warranty clause for a century.

So copyright law already accepts that authorship is not the same as typing, and the industry accepts it too, as long as the fingers on the keyboard belong to a human being who has been paid and asked to stay quiet.

Which means the objection to AI is not really about who did the writing, it is about who deserves the status.

Which is it

Here is the asymmetry I cannot get past.

When a language model produces something good, the industry position is that no human authored it, since there is nothing to copyright and nobody to credit, and no art was involved anywhere in the process. When a language model produces something defamatory or plagiarised, the position flips instantly: the human who prompted it is responsible, is liable, should have checked. Publishers ask you to warrant that the text is your own creation and simultaneously insist that it can’t be.

You cannot have both, and the choice is not subtle. Either the person steering the machine is the origin of the work, in which case they own it and answer for it, or they aren’t, in which case go find someone else to sue. Every legal system that has looked at this has quietly chosen responsibility without granting credit, and that isn’t a principle, it’s a convenience.

She is right about AI editors

I want to give her the strongest point she makes, because I ran into it myself and it cost me months.

Ask a model to edit your manuscript and it will edit, which is the whole problem. It has no way to say “this chapter is fine, leave it alone,” so it manufactures suggestions to justify its own existence, and because editorial advice is not verifiable, and you cannot check whether merging two characters would in fact be better, you have no way to tell useful notes from generated ones. Fenton describes it as sounding like an answer while being nonsense, and that matches what I saw.

What she gets wrong is the conclusion, because this is a tooling failure, not a property of the technology, and it is fixable.

What worked for me was refusing to ask the broad question. “Edit this chapter” gets you the same thing as “write me a book,” so I went and read the manuals instead, three of them, and not books about how to write: Susan Bell’s The Artful Edit, Dreyer’s English, and Amy Einsohn’s The Copyeditor’s Handbook. I read them the way an engineer reads API documentation, looking for nameable operations instead of taste. What came out of that is the reason human editors price developmental work and copyediting differently: they are separate layers that ask separate questions, and running them at the same time is why “edit my manuscript” fails.

So I split it into three passes, each one finished before the next begins. Pass one works a single scene at a time on pacing, interiority, dialogue subtext and whether the emotional beat on the page is the one I think is there. Copyediting comes second on that same scene, the four Cs plus an adverb audit and a genre-cliché list I had to build myself for military sci-fi. Only the third pass runs across the whole manuscript, once, and does nothing but hunt repeated images: every sensory detail or physical tic used three or more times to signal the same thing, logged, then replaced from the third instance on. That last one catches what I stopped being able to see, because I had been inside the book far too long and the model had not.

I wrote the whole method up separately, and the three prompts are available to paid subscribers here.

None of this removes me from the process, which is the point Fenton and I are actually arguing about. The passes annotate and flag and propose, and I still decide whether any given replacement is better than what I wrote, and after a few hundred inline notes you start catching your own patterns in the next draft before the model does.

That is a long way from a human developmental editor, and it is also a long way from nothing, which is what most authors can actually afford, and Fenton says so herself: she’d love a more cost-effective way for authors to get editorial input. She started her Substack partly for that reason. I just think she has decided in advance that the cost-effective way cannot be this one.

The people who can’t write

The line that made me put the coffee down: “if you do not have the writing skill to write a book, then perhaps you do not get to write one.”

Storytelling is older than writing by something like thirty thousand years. Every oral tradition that ever existed was built by people with no letters at all, and we only have Homer because somebody eventually wrote down what a non-writer composed. Skill with sentences is one delivery mechanism for narrative, and it happens to be the one that publishing sells, which is why an agent would mistake it for the whole art.

I know people whose imagination I would trade mine for and whose sentences are a mess, some of them dyslexic, some working in a third language, some simply never taught. Under Fenton’s rule they don’t get to have their story exist, and the world is not better for that, it’s just quieter in a way that happens to protect the people who already have the skill.

And the tools argument cuts deeper than she allows, since spellcheck is a shortcut, Grammarly is a shortcut, Scrivener’s structure view is a shortcut, and a copyeditor who silently fixes your comma splices for eighteen months is a very expensive shortcut. Nobody has ever said a dyslexic writer using dictation software didn’t really write their book. The line always seems to fall just after whatever the person making the argument already uses.

The random number

Her best rhetorical moment is a quote from Claude’s own documentation, where the model explains that when either “overcast” or “grey” would do, the choice comes down to a random number. She wants authors who care intensely which word it is.

So do I, and that is the entire argument for keeping a human in this, and she’s presented it as an argument against.

A model with nobody watching produces exactly the slop she is describing, and there are hundreds of thousands of those books on Amazon right now, all averaging toward the same texture, so the difference between that and something worth reading is a person going through it deciding that no, it’s not grey, it’s the colour of the sky over a firebase at 0500, and that this paragraph is doing nothing and comes out. Word by word, decision by decision, the same job she’s describing, just starting from a draft instead of a blank page.

You can tell the difference in the reading, which is why I’m not worried about the flood in the long run, since readers do not owe anyone patience.

What I’m actually asking for

I don’t want publishers to accept unedited machine output, and I don’t think they will or should. What I want, and it feels like a small ask, is for the argument to stop pretending it is about the writing when it is about gatekeeping, because those need different answers.

If a publisher’s policy says do not upload any part of your manuscript to any LLM – and Fenton reports that some do, which she flags herself as extraordinarily broad – then that policy is not protecting the quality of anyone’s sentences. It is protecting a screening function, and the screening function is worth understanding. Traditional publishing’s service to readers was deciding what was good enough to print, and its service to authors was access to shelf space, the second of which self-publishing dissolved fifteen years ago. Tools that let unskilled storytellers produce readable manuscripts threaten the first, and the response has been to add a new admission requirement at precisely the moment the old one stopped working.

Maybe that’s the wrong read and there is a version of the AI policy that really is about the writing. If someone can show me a publishing house that will take a manuscript from a dyslexic storyteller who dictated and machine-drafted and then spent two years rewriting every line by hand, I will drop the whole argument. I have not found one, though I admit I have been looking from the outside, and I don’t think the question has been put to anyone seriously.

The daughter still can’t ride a bike, and there’s a trailer in the garage.