In 2022, reporter Will Douglas Heaven at MIT Technology Review submitted a piece about large language models to an internal review workflow that used an early GPT-4 preview for structural feedback. The system flagged that his third section buried the most newsworthy claim — a genuine editorial observation that a human sub-editor had missed on first pass. The story ran with that section moved to the lede. Heaven later wrote about the experience, noting the AI had caught a structural problem, not a grammatical one.
Most writers think of editing as one thing. Professionals divide it into at least three distinct operations, each requiring a different cognitive lens.
Proofreading operates at the surface: spelling, punctuation, typographical errors, consistent hyphenation. Rule-based. Deterministic. This is where spell-checkers have worked since the 1970s.
Copy-editing goes one level deeper: grammar, sentence clarity, word choice, factual consistency, style-guide compliance. Still largely rule-governed, but judgment enters.
Structural (developmental) editing is the deepest layer: Does the argument hold? Is the pacing right? Does the reader know what they need to know when they need to know it? This requires holding the entire work in mind simultaneously — something humans find hard and AI finds… different.
AI language models are genuinely strong at copy-editing. They have been trained on enormous corpora of edited prose and can flag passive-voice overuse, nominalisation, hedging language, and inconsistent tense with speed no human matches at scale.
Where AI surprises most writers is at the structural layer — not because it reasons like a developmental editor, but because it can hold a 5,000-word draft in a single context window and compare paragraph 12 against paragraph 2 without fatigue or the anchoring bias humans develop after reading a draft three times.
In 2023, the literary agency Janklow & Nesbit began experimenting with Claude for preliminary manuscript notes. Agents reported that the tool consistently identified chapters where a character's motivation was inconsistently established — catches that previously required a second human reader.
AI cannot evaluate whether a piece is true in the journalistic sense, whether a scene is emotionally authentic to a real human experience, or whether a stylistic choice is daring versus merely strange. Those remain human editorial judgments.
You give AI the draft and ask: "What are the three biggest problems with this piece?" No fixes yet — just diagnosis. Forces the model into a structural frame rather than a line-edit frame.
You give AI specific criteria — "flag every sentence over 30 words," "identify every claim without evidence," "find where the reader loses the thread" — and it audits against those rules.
You give AI two versions of a passage and ask which better serves a stated goal. Useful when you're caught between revisions and need a reader reaction, not a rule.
You give AI a weak passage and ask it to demonstrate an alternative — not to replace your prose, but to show you what a different approach might look like so you can decide.
The journalist and author Steven Johnson, writing in The New York Times Magazine in 2023, described using AI in diagnostic mode on his book drafts: "I stopped asking it to fix things. I started asking it to tell me what was broken. That's a completely different conversation."
Paste a short passage (3–8 sentences) of any draft writing — fiction, journalism, essay, anything. Then ask the AI editor to diagnose its structural problems before suggesting any rewrites. Practice separating diagnosis from prescription.
Try prompts like: "What are the two biggest structural problems with this passage?" or "Where does a reader lose the thread here?"
In 2023, Joanna Maciejewska, a fantasy novelist, published detailed notes on her Substack about using Claude to map tension curves in her manuscript. She asked the model to rate each chapter's tension on a 1–10 scale and explain its reasoning. The resulting chart revealed that three consecutive middle chapters all scored 3–4 — a sag she had felt but could not precisely locate. She later called it "the most useful single piece of editing feedback I got on that book."
Pacing is not synonymous with speed. A slow scene can have intense forward momentum; a fast-action sequence can feel oddly inert. Pacing is the reader's felt sense that something is at stake and changing — that the story is earning their attention moment to moment.
The craft elements that generate pacing include: sentence length variation, scene-to-summary ratio, the density of new information per paragraph, unresolved questions (micro-tension), and the distance between a dramatic question being posed and answered.
AI does not feel boredom. But it can identify proxies for the conditions that produce boredom in readers. Specifically:
Information density drops: When paragraphs contain mostly re-statement of what was already known, rather than new facts, questions, or complications.
Sentence length homogeneity: Prose that runs at a consistent medium length for many consecutive sentences tends to lull — the rhythm stops surprising.
Scene without consequence: Extended scenes where no character decision or revelation changes the situation. The reader senses the story has paused to describe rather than to act.
Unresolved question drought: Strong pacing keeps open questions in the air simultaneously. When all questions from a prior scene are answered and no new ones are raised, momentum stalls.
A 2023 analysis by researchers at the Allen Institute for AI (AI2) found that GPT-4 could identify "tension arcs" in short fiction at roughly 74% agreement with trained human editors when prompted with structured evaluation criteria. Without structured prompts, agreement dropped to 41% — a useful reminder that prompt quality determines output quality.
Zadie Smith, in a 2023 interview with The Guardian, noted that some of her favourite passages in her own work would score poorly on any tension metric — they are deliberately ruminative. The point of tension mapping is not to eliminate slow passages but to ensure your slow passages are intentional choices, not invisible problems.
Paste a scene or passage (at least a paragraph) and ask the AI to score its tension 1–10 and explain the specific factors driving that score. Then experiment with what the AI suggests would raise or lower the score.
Try: "Rate this passage's pacing 1–10 and name the two specific things most responsible for that score."
In October 2023, literary magazine Granta published an editorial by their digital team documenting an internal experiment: they ran 30 submissions through GPT-4 for a "clarity pass" before human editorial review. When the human editors were shown the AI-touched manuscripts without being told which had been processed, they rated the AI-edited prose as less distinctive on average — noting a flattening of idiosyncrasy, unusual syntax, and deliberate rhythm breaks. Granta's conclusion: AI copy-editing, applied without constraint, optimises for readability at the cost of voice.
Voice is not personality sprinkled on top of prose. It is a set of specific, consistent micro-decisions: sentence rhythm, the ratio of concrete to abstract, when the writer interrupts themselves, how they handle transitions, what they never explain, what they over-explain, the particular distance they maintain from their subject.
AI models, particularly when instructed to "improve" or "clarify," default toward consensus patterns in their training data — which is to say, toward what most edited prose looks like. That means longer sentences get broken, unusual sentence openings get normalised, and the choices that make a writer recognisable get quietly removed.
The solution is to explicitly teach the model your voice before asking it to edit. This has three practical steps:
Step 1 — Voice sample: Provide 3–5 paragraphs you consider to be strong examples of your best writing. Tell the AI: "This is my voice at its clearest. Study the sentence rhythm, the way I handle transitions, the level of formality."
Step 2 — Voice inventory: Ask the AI to list the specific stylistic features it observes in your sample. Review its list. Add anything it missed; correct anything wrong.
Step 3 — Constrained edit: Only now submit the passage needing editing, with the instruction: "Edit for [specific issue] without changing any of the voice features we identified."
In 2023, essayist and critic Hanif Abdurraqib described his use of AI editing tools in an interview with Vulture: "I use it like I'd use a cold read from someone who doesn't know me yet. I give it a lot of context about what I'm trying to do. Without that context, it tries to make everything sound like everything else."
Sentence fragments used deliberately. Unusual punctuation rhythm. Unexpected tonal shifts. Repetition used as a device. Sentences that break grammar rules for effect. Any construction that appears repeatedly — it is likely a signature.
Accidental repetition (a word used three times in one paragraph without intent). Unclear antecedents. Tense inconsistencies. Paragraphs that could be read two ways but you only intend one. Real typos, not stylistic ones.
The best human editors have always worked by this principle: the job is to help the writer be more themselves, not to impose an external standard. That principle applies equally to AI editorial work. The difference is that a human editor unconsciously absorbs your voice after reading your work. AI requires you to make that context explicit.
Use the voice-lock technique: first share a strong writing sample and ask the AI to identify your stylistic signatures. Then submit a different passage for editing with those signatures explicitly protected.
Start with: "Here is a writing sample I'm proud of. Please identify the specific stylistic features that define my voice."
In February 2024, The Atlantic published a reported essay by Ian Bogost examining the phenomenon of writers becoming trapped in endless AI revision cycles. Bogost interviewed a cohort of freelance journalists and found that several had used AI editing tools to revise pieces 15–20 times before filing — compared to 3–5 passes previously. The pieces were not measurably better; in two documented cases, editors at the publications noted that the final filed drafts felt over-worked — the prose had lost spontaneity. The problem was not AI but the removal of any forcing function to stop.
AI-assisted revision introduces a specific new failure mode: because revision is now nearly frictionless, writers can revise indefinitely. Every pass surfaces new suggestions. Every suggestion, implemented, creates new micro-inconsistencies that the next pass finds. The draft is always improvable. The loop never ends.
Traditional editorial workflow had natural stopping points built in: a deadline from an editor, the physical cost of retyping, the scarcity of trusted readers. AI removes all three. Without a replacement structure, writers drift.
In 2023, the content team at Mailchimp published internal guidelines on AI-assisted editing that included a mandatory "revision ceiling" — no more than four AI-assisted passes per piece. Team leads reported a 30% reduction in the time-to-publish for long-form content without editors reporting quality degradation. The ceiling was the key variable.
AI cannot tell you a piece is finished. It can always find something to refine. The judgment that a piece has achieved its purpose is a human editorial call, and a crucial one.
Signals that a draft is done, despite AI suggestions to the contrary: the piece does what you set out to do; every remaining AI suggestion addresses a feature you want rather than need to change; you have implemented all changes you can justify; you would not be embarrassed to have the work published in its current state.
The novelist Anne Lamott wrote in Bird by Bird (1994) about the concept of "good enough" — the moment when further revision improves the piece technically but no longer brings it closer to its essential intention. That principle predates AI but applies with new urgency to AI-assisted revision workflows.
AI is an instrument for serving your editorial vision. When you find yourself implementing AI suggestions without being able to articulate why they improve the piece toward your stated goals, the tool has inverted the relationship — and it is time to stop.
Write a "done criteria" statement for a piece you're working on, then run one focused revision pass with a specific goal. Practice declining AI suggestions that don't serve your stated criteria.
Start with: "Here are my done criteria for this piece: [your criteria]. Now review this passage for [one specific issue only] and ignore everything else."