Writing With AI as a Tool

I edit fiction for a living. I’ve read manuscripts sentence by sentence, argued with writers about a single comma, and spent whole afternoons on whether a character would say “okay” or “fine.” So when people find out I don’t treat AI as an enemy in that process, they usually want an explanation. Fair enough.

Here’s the case, plainly.

The objection, stated honestly

The strongest version of the anti-AI argument isn’t “it’s lazy” or “real writers don’t need help.” It’s this: writing is thinking, and if you outsource the sentence-making, you outsource the thinking that produces it. A tool that generates prose for you doesn’t just save time — it removes the friction where understanding actually happens. Writers who lean on AI to draft, the argument goes, end up with pages that read fine and mean nothing, because nothing was worked out in the writing of them.

I take that seriously. It’s true of a certain use of these tools. It is not true of every use, and the difference matters more than most of this debate admits.

Tool, not author

A hammer doesn’t build the house. It drives the nail the carpenter decided needed driving, at the angle the carpenter chose, because the carpenter understands load and grain and what happens if the joint fails. Nobody worries the hammer is doing the thinking.

AI used well in writing works the same way, if you use it that way. I use it to stress-test an argument before I commit to it, to generate ten bad title options so I can see why the eleventh one is right, to catch a structural problem I’m too close to see in my own draft. None of that replaces the sentence-level decision-making that is the actual craft. It clears space for me to do more of it, faster, on the parts that matter.

The failure mode the critics describe — prose that reads fine and means nothing — happens when a writer asks the tool to make the decisions instead of asking it to help test the decisions they’ve already made. That’s a discipline problem, not a tool problem. Writers have been outsourcing thought badly for as long as there have been shortcuts to take. Formula fiction, ghostwriting without a strong brief, workshop groupthink — all of it produces competent, empty prose, and none of it required a language model.

The tell is in the editing

Here’s how you know the difference in practice. Work that was drafted by having AI do the thinking reads smooth and dies under any real pressure — ask why a character made a choice, or what the scene is actually doing structurally, and there’s no answer, because no one worked it out. Work where AI was a tool in a writer’s hand holds up under that pressure, because the writer still owns every decision; they just got there with better tools.

I can tell within a page which manuscript I’m holding. So can any editor who reads for a living. The objection that AI-assisted writing is inherently hollow assumes the writer wasn’t paying attention. Some weren’t. Most of my clients are.

The em dash accusation

There’s a smaller, louder version of this argument that isn’t really about thinking at all. It’s about tells — the em dash, certain rhythms, a fondness for “moreover” or three-item lists that snap shut with a rule of three. Add to the list: “delve,” “tapestry,” “boundaries,” “navigate,” “landscape,” “underscores,” “in today’s world,” “it’s important to note.” Someone reads a piece, spots one of these, and declares it AI-written, case closed.

This one I have less patience for, because it isn’t a craft argument. It’s pattern-matching dressed up as one. None of these words are new, and none of them are wrong. “Delve” is a perfectly good verb that predates the transformer by about a thousand years. “Navigate” is the natural word for moving through something complicated. Writers reached for “tapestry” as a metaphor long before any model was trained on one. The em dash didn’t start with AI either — Dickinson ran on it, Didion built whole essays out of it, and plenty of careful human writers reach for it because it does something a comma or period can’t. Treating a word or a punctuation mark as proof of authorship gets the causality backward: language models learned that vocabulary and rhythm from human writing, not the other way around. Punishing writers for using it now is punishing them for sounding like the thing the tool was trained to imitate.

It’s also just a bad test. It has a high false-positive rate against any writer with a formal register or a broad vocabulary, and a high false-negative rate against anyone who ran their AI draft through even one pass of real revision, which strips those words out fast. A test that flags Didion for the em dash and a well-read grad student for “delve,” but misses the sloppy AI draft that got a light edit, isn’t measuring what people think it’s measuring. It’s measuring surface familiarity with a style, not origin.

If the concern is really about origin and disclosure, ask about that directly. Judging authorship by word choice is a shortcut that flatters the accuser more than it informs anyone.

What this isn’t

This isn’t an argument that AI-generated prose should pass as a human’s finished work without acknowledgment, or that publishers and readers don’t have a right to know how something was made. Those are real questions, and I don’t think “it’s just a tool” settles them by itself. Disclosure and authorship norms are worth arguing about honestly, separate from the craft question.

It’s also not an argument that every writer should use these tools. Plenty of good writers won’t touch them, and their work is no worse for it. The tool doesn’t make the writer. It never did.

Where I land

I use these tools every week and I judge manuscripts exactly as I did before. Those two facts aren’t in tension, and the reason they aren’t is the only thing I’d want a writer to take from this: the tool never determined the value of the work. Attention did.

A writer who is paying attention uses a language model the way they use a thesaurus, a reader they trust, or a long walk — as something that helps them arrive at a decision they were already circling. A writer who isn’t will produce something competent and empty, and would have found a way to do that in 1985.

So I don’t think “did you use AI” is a useful question to ask of a manuscript. “Who made the decisions” is. It’s answerable, it predicts whether the book will hold up under pressure, and it’s the same question I’d have asked about a ghostwriter, a heavy-handed development editor, or a workshop that sanded a voice flat.

What I’d ask of writers is honesty about the answer — with readers where it matters, and with themselves always. The first is a disclosure question and deserves a real argument, which it hasn’t had yet. The second isn’t up for debate. You know whether you did the work. The page tells everyone else soon enough.