When novelist Cormac McCarthy submitted the manuscript of The Passenger to his editor at Knopf, the text was riddled with sentences that any AI grammar tool would flag: comma splices used as deliberate rhythm, dialogue stripped of quotation marks, subordinate clauses piled seven deep. McCarthy's refusal of conventional punctuation is not error — it is style. Grammarly, run on a passage from the finished book in a 2022 test by The Atlantic, returned 47 alerts on a single page. Every one of them missed the point.
The incident is instructive not because AI tools are worthless, but because understanding what they are trained to optimize for is the prerequisite for using them wisely.
Modern AI writing assistants — Grammarly, ProWritingAid, the grammar layer inside Microsoft Editor — are trained primarily on large corpora of edited, published prose: news articles, academic papers, business documents. The model learns statistical patterns: what word sequences appear most often in "clean" text versus text that was later revised. That training signal is excellent for catching genuine errors in functional writing. It is a poor guide to literary style.
The key concept is perplexity. A language model assigns lower perplexity to sequences it has seen often. An unusual syntactic move — an intentional fragment, an inverted sentence, a Latinate periodic construction — registers as high-perplexity and gets flagged. The model cannot distinguish between "I made a mistake" and "I made a deliberate artistic choice that departs from statistical norms."
AI grammar tools optimize for statistical normalcy, not correctness in the absolute sense. They flag departures from the center of their training distribution. Literary voice often lives at the edges of that distribution — and that is its power.
Despite the caveats, AI editing tools provide genuine value for specific categories of error. Homophones (their/there/they're), subject-verb disagreement caused by intervening clauses, doubled words ("the the"), and missing closing punctuation are all cases where the model's statistical training aligns well with the writer's intent. A 2021 study by the University of Pittsburgh's Writing Center found that Grammarly caught 72% of grammar errors in undergraduate essays that human peer reviewers missed — a substantial benefit for functional prose.
The strategic writer uses these tools on a second pass, after the stylistic decisions are locked. You run the checker not to learn how to write, but to catch the typos you stopped seeing after your fifteenth read-through. Then you selectively override.
Many AI tools offer aggregate scores: Grammarly's "Correctness," ProWritingAid's "Readability," Hemingway App's grade-level rating. These scores carry an implicit claim — that writing closer to Grade 6 reading level is inherently better. This is a design choice, not a linguistic truth.
Henry James averages a Hemingway App score of "PostGraduate" with most sentences flagged red. Toni Morrison's Beloved returns a readability grade of 10+ and hundreds of passive-voice flags. The score tells you how far a text deviates from a particular norm of plain style. It tells you nothing about whether that deviation serves the work.
The practical discipline is this: know what each metric actually measures before deciding whether to care about it. "Passive voice" flags catch genuine cases of agency-burying in journalism; they misfire constantly in fiction where the passive is chosen to withhold information or weight a phrase differently.
Before running any AI editing tool, write one sentence describing the stylistic choices that define this piece — the moves you will protect. Keep it visible while reviewing suggestions. Every flag is a question: does accepting this suggestion protect or erase a deliberate move?
Paste a short passage of your own writing (or one of the examples below) into the chat. Ask the AI to act as an AI grammar tool and flag every "issue" it sees. Then push back: ask it to identify which flags represent genuine errors versus intentional style choices. Explore how it distinguishes — or fails to distinguish — between the two.
In early 2023, journalist Casey Newton wrote in his Platformer newsletter about using Claude to read a draft of a long-form technology profile — roughly 6,000 words — and return a structural memo: where does the argument go slack, which anecdotes earn their length, where does the reader lose the thread. Newton reported that Claude's structural notes were "startlingly accurate" on two of three major issues his human editor later identified, and wrong in a third case in a way that was itself illuminating — the AI flagged a digression that turned out to be the emotional heart of the piece.
That case encapsulates the structural editing relationship: AI as a fast first-pass reader whose responses reveal something real about the text, even when the diagnosis is wrong.
Structural editing — sometimes called developmental or substantive editing — operates above the sentence. It asks: Does this piece have a clear purpose? Does the order of sections serve that purpose? Are transitions doing work? Is the pacing matched to the stakes? In traditional publishing, structural editing is the most expensive editorial service because it requires a reader who can hold the whole piece in mind.
AI systems can approximate this function because large language models process text in large context windows — GPT-4 Turbo and Claude 3.5 Sonnet both handle 100,000+ tokens, roughly 75,000 words. They can "read" a full draft in one pass and return a coherent response about the whole. This is genuinely new. No human beta reader reliably provides structural notes on a 10,000-word draft within minutes.
The quality of structural feedback depends almost entirely on how you prompt. Vague prompts produce vague feedback. The following prompt types have been documented as effective by writing instructors and authors who have published their workflows:
The Reader Experience Prompt: "Read this draft as an intelligent general reader encountering it for the first time. At the end of each section, describe in one sentence what you now understand to be the piece's central claim, and whether that understanding has shifted since the previous section."
The Tension Map Prompt: "Identify the five moments in this draft where a reader is most likely to either accelerate (lean in) or decelerate (skim). For each, explain what causes that response."
The Missing Thread Prompt: "What question does this piece raise in the first 300 words that it never fully answers? Is that omission intentional?"
Each of these prompts forces the AI to make a specific, falsifiable claim rather than offering vague encouragement. You can then disagree, probe, or redirect.
A structural editing prompt is well-designed when it could produce an answer you disagree with. If every possible answer would feel useful and non-threatening, the prompt is too vague to teach you anything.
AI structural editing has documented failure modes. The most significant: it tends to prefer conventional argument structures. A piece that buries the lede intentionally, or that builds toward a conclusion the reader must construct, or that uses juxtaposition rather than transition — these moves often get flagged as problems when they are features. Editor Sari Botton, writing in Catapult in 2023, noted that when she ran experimental personal essays through AI structural analysis, the AI consistently recommended moving the "main point" earlier. "It was right for most essays," she wrote. "It would have killed the specific essays I value most."
Additionally, AI structural notes treat every draft as if it exists in isolation. They have no memory of your stated intentions from a previous conversation unless you re-supply them. Always open a structural editing session by telling the AI what this piece is trying to do, for whom, and in what form it will be published. Context changes everything about what counts as a structural flaw.
Before pasting a draft for structural review, write two sentences: (1) what this piece is trying to do, and (2) what structural move it makes that a conventional reader might misread as a flaw. Paste those sentences as your opening instruction. The AI's subsequent notes will be calibrated to your actual goals.
Use this chat to practice structural editing prompts. You can paste a draft (or outline), or work through a hypothetical structure. Start by stating the piece's purpose and any structural move you're protecting. Then use one of the three prompt types from Lesson 2: Reader Experience, Tension Map, or Missing Thread.
In 2023, researchers at the University of Pennsylvania published a study in PLOS ONE analyzing 1 million academic abstracts before and after the widespread availability of ChatGPT. They found a statistically significant increase in certain high-frequency words — intricate, commendable, meticulous, pivotal — across abstracts published in 2023 compared to 2021. The words themselves are unremarkable; their sudden ubiquity revealed a shared upstream source: large language models trained on similar corpora, producing similar lexical choices when asked to "improve" academic prose.
The finding does not prove AI editing is harmful. It proves that AI editing, used without critical oversight, produces convergence — a narrowing of the vocabulary and sentence patterns that a field uses to express itself. The same risk applies to literary prose.
When a language model is asked to "improve" a sentence, it draws on the center of its training distribution. That center is competent, clear, moderately formal, and largely anonymous. It is the prose of a well-edited magazine article — never embarrassing, never distinctive. If you accept its suggestions uncritically across a whole manuscript, you end up with prose that is technically better by standard metrics and considerably less yours.
Literary scholars Naomi Baron and Mark Riedl have both written about this risk: AI writing assistance, at scale, may reduce the variance of published prose over time, making the literary landscape more homogeneous. This is a collective problem, not just an individual one.
The solution is not to avoid AI line editing but to prompt it differently. Instead of "improve this sentence," use prompts that foreground your voice as a constraint the AI must work within:
"This sentence is unclear but needs to retain its staccato rhythm and its refusal to explain. Suggest three alternatives that fix the clarity problem without changing those features."
"I use semicolons to create a pause that is heavier than a comma but lighter than a period. Suggest a revision of this paragraph that keeps exactly that weight — do not substitute dashes or conjunctions."
"The word choice here feels weak but the specific quality I'm after is flatness, not vividness. Suggest three alternatives that are flatter and more exact, not more colorful."
Each prompt gives the AI a problem to solve (clarity, precision) while specifying which stylistic features are off-limits for revision. This is the difference between using AI as an authority and using it as a constrained tool.
The appearance of words like meticulous and pivotal in academic writing post-ChatGPT is a measurable signal of what happens when AI suggestions are accepted without interrogation. Watch your own editing sessions for words you would never have chosen but accepted because the AI offered them fluently.
Writers who have developed effective AI-assisted line-editing workflows typically describe some version of a three-pass structure. Pass 1: Read the draft alone and underline every sentence that feels wrong — unclear, weak, or structurally awkward. Identify why in a margin note. Pass 2: For each underlined sentence, write a prompt that names the specific problem and the stylistic constraints the revision must honor. Run the AI on each sentence individually. Pass 3: Read the revised draft aloud. Any sentence that sounds unlike you goes back — either to the original version or to a further iteration with a better-calibrated prompt.
The final pass is the one most writers skip and most need. Reading aloud is the fastest test of whether a revision sounds like your voice or like competent-anonymous prose. The ear is harder to fool than the eye.
After an AI-assisted editing session, run a vocabulary audit: search your manuscript for words you flagged as AI suggestions and accepted. Ask: would I have reached for this word on my own? If not, replace it with the word you would have chosen — even if the AI's suggestion is technically stronger. Voice is an accumulation of specific choices, not a general quality.
Choose one or two sentences from your own writing that feel weak or unclear. Describe the problem precisely. Then describe the stylistic feature — rhythm, register, tone, syntax pattern — that must be preserved. Write a constrained revision prompt and submit it. When the AI responds with alternatives, push back: ask it to try again if any suggestion sounds anonymous or generic.
In October 2023, novelist Walter Kirn published an essay in The Atlantic describing his experimental use of AI throughout the revision of a long-form magazine piece. Kirn did not use AI to draft; he used it at three specific revision stages: first to read his structural outline before writing began, second to identify pacing problems in a full first draft, and third to flag clichés in his final polish pass. He retained all authority over actual changes. "The machine," he wrote, "is a mirror I can argue with. That's different from a mirror I simply accept."
Kirn's framework — AI as an interlocutor rather than an authority — describes the disposition that makes AI-assisted editing productive rather than corrosive.
An effective AI-integrated revision workflow operates in distinct layers, each with a different tool set and a different type of question being asked. The layers should not collapse into one another — running a grammar check before structural problems are resolved is a waste; running a structural analysis after you've finalized the line edit may uncover problems you no longer have bandwidth to fix.
Layer 1 — Structural (Draft Complete): Use AI to read the full draft for architecture. Use falsifiable prompts (Tension Map, Missing Thread). Identify problems you will address in the next draft. Do not touch individual sentences yet.
Layer 2 — Line Edit (Structure Locked): Work sentence by sentence on the passages you've identified as weak. Use constrained revision prompts. Run the read-aloud test after each session. Track AI-originated vocabulary and audit it before moving on.
Layer 3 — Copy Edit (Voice Locked): Run grammar and style tools. Accept only error corrections. Override anything that touches rhythm, register, or word choice you have deliberately established in Layer 2.
The most common mistake in AI-assisted revision is collapsing the layers — accepting a grammar tool's word-choice suggestion during structural editing, or doing line edits before structure is resolved. Each layer should feel complete before the next begins.
As AI-assisted editing becomes standard, a practical question arises: how do you maintain a clear record of what you wrote versus what AI suggested? Several professional practices have emerged. Version control: save a pre-AI draft with a date stamp so you always have the unmediated version. Session logs: keep a text file of AI prompts you used and what category of change they produced. This is useful both for your own clarity and, increasingly, for disclosure requirements from publishers.
In 2023, major publishers including Penguin Random House and HarperCollins began including AI-use disclosure clauses in author contracts. These clauses typically ask authors to disclose if AI was used to generate substantial portions of the text — not if it was used as an editing tool. The distinction matters, and maintaining records supports your ability to make it clearly.
Walter Kirn's phrase — "a mirror I can argue with" — captures the productive disposition. This means: read every AI suggestion as a question about your work, not an instruction. "This paragraph loses momentum" is not a fact; it is a reading, and readings can be wrong. Your job is to decide whether the AI's reading is accurate about your readers' likely experience, and whether addressing the identified problem serves the piece's actual goals.
The authors who report the most productive AI editing relationships are those who maintain strong prior commitments — clear positions on what a piece is doing and why — before they show it to any AI. Writers who enter revision uncertain of their own intentions tend to accept AI suggestions more readily, because any direction feels like progress. The AI fills an editorial vacuum that should be filled by the writer's own vision first.
Before your next revision session, write three sentences: (1) what this piece is trying to do, (2) which AI tools you will use at each revision layer, and (3) one stylistic feature you will not allow any AI suggestion to change. This "editing manifesto" takes two minutes and preserves the conditions under which productive AI-assisted editing is possible.
In this lab, you'll practice the full interlocutor stance: describe a piece you're revising, state your editing manifesto (what the piece does, what layers you'll use AI for, what one feature you'll protect), and then run the AI through one full revision layer — structural, line-level, or copy edit — using the appropriate prompt type. Push back on any suggestion you disagree with. Practice the "argue with the mirror" disposition explicitly.