Every few weeks, another writer gets caught using AI and reaches for the same shield: I used it like an editor. Like a brainstorming partner. It surfaced in The Atlantic's coverage of authors quietly leaning on ChatGPT, in the reaction to Olga Tokarczuk's remarks about AI and literature, and again in the Lauren Gilgan controversy around AI-assisted work in the New York Times orbit. The defense is consistent enough to deserve examination on its merits. At YouWrite we build tools like Refine that sit near this exact line, so we have skin in getting the distinction right.
Let me steelman the defense first, because it isn't nonsense.
The honest case for the analogy
Writers have always used other minds. Maxwell Perkins reshaped Thomas Wolfe's manuscripts so aggressively that scholars still argue about authorship. Gordon Lish cut Raymond Carver's stories down to the bone, and we still call them Carver's stories. Every serious novelist has a first reader, a spouse, an agent, a workshop. The romantic image of the solitary writer is largely a marketing invention.
So when a writer says an AI helped them think through a stuck chapter, that maps onto something real. Talking to a competent interlocutor about structure genuinely helps. Getting a bulleted response to what are five ways this scene could end is not categorically different from asking a friend the same question at dinner. If we allow the friend, the categorical objection to the machine collapses.
That's the strong version. Now the problems.
Where the analogy quietly breaks
An editor responds. A model generates.
Maxwell Perkins could not write Look Homeward, Angel for Wolfe. He could only respond to what Wolfe put in front of him. The asymmetry was built into the relationship: the writer produced, the editor reacted.
A language model has no such constraint. You can ask it to draft the chapter first, then pretend you're editing its output. The direction of creative flow reverses without anyone noticing, including sometimes the writer. You end up with prose you approved rather than prose you generated. Approval is not authorship, or at least not the same kind.
An editor has no authorial stake. A model produces adoptable prose.
When a human editor suggests a line, the suggestion arrives marked. It comes in a margin note, a track-changes bubble, a phone call. The writer has to consciously decide to incorporate someone else's language, and that friction is the whole point. It forces attribution to stay legible, at least to the writer.
Model output arrives unmarked. You can copy a sentence straight into your draft, and by the third revision you genuinely cannot remember whether you wrote it or accepted it. The friction that made borrowing visible is gone.
A brainstorming partner remembers. A model doesn't care.
A real brainstorming partner has continuity with your project. They remember the argument you were having about the protagonist's mother three weeks ago. They push back when your new idea contradicts something you told them in October. That memory creates accountability to the work's own logic.
A model, even one with long context, has no investment in your creative continuity. It will happily help you contradict yourself. It will validate a bad direction as enthusiastically as a good one. The pushback of a real collaborator, the wait, you already tried this, is structurally absent.
A practical test you can actually apply
Here is a question worth asking about any passage in your draft:
Did this language originate in my head, or in the model's?
Not did I approve it. Not does it match my voice. Origination. If you cannot honestly say the sentence formed in your mind before it appeared on the page, it is model prose you are wearing. That may still be fine for a memo, a product description, a LinkedIn post. It is not fine for work you present as authored.
The test has useful gray zones:
- You describe a scene to the model in your own words, it suggests reordering two paragraphs, you agree. Origination is yours. Structural feedback sits closer to genuine editing.
- You give the model a bulleted outline, it writes three paragraphs of prose, you tweak adjectives. Origination is the model's. You are the editor now, not the author.
- You draft a paragraph, ask the model why it feels flat, it points to a buried verb, you rewrite. Origination stayed with you. This is the strongest version of the editor analogy holding.
- You are stuck, you ask for ten opening lines, you pick one and build from it. Origination is split, and honesty requires you to notice that.
The line is blurry for outlining, diagnosis, and structural response. It is clear for prose generation. Pretending otherwise is where the defense turns into cover.
Where YouWrite sits, honestly
Our Refine tool works on text you already wrote, tightening rhythm and cutting filler. That places it closer to the editor end of the spectrum. But any tool that can rewrite a sentence can also generate one, and users decide where the line falls. We can nudge toward genuine revision through interface design. We cannot enforce origination. No one can. That is the writer's problem, and it stays the writer's problem.
The competitor framing, and where it's too clean
A cluster of AI writing coaches, Nicolas Cole's Ship 30 for 30 adjacent material, Sudowrite's marketing, various Substack advocates, all use the language of AI as collaborator or AI as thought partner. The framing is not dishonest. It is strategically vague. It elides the origination question by focusing on the writer's felt experience of the process. If it felt collaborative, it was collaboration.
Felt experience is not the standard readers apply when they find out. Readers care about origination. The Gilgan reaction wasn't about whether she experienced AI as a partner. It was about whose sentences those were. The collaborator framing works right up until the audience learns the split and decides for themselves.
What to do on Monday morning
Keep a log for one week. Every time you use a model on a piece of writing, note whether you asked it to respond to your words or to produce words. Tally the ratio. If produce outnumbers respond, the editor analogy is not describing what you're doing, and you owe yourself a more accurate word for it.