Hand AI a paragraph and say only "polish this up," and two problems tend to follow: the result reads smoothly but doesn't sound like you, and you can't tell exactly what changed — so you either accept the whole thing or scrap it wholesale. The root cause: one vague instruction actually carries three separate jobs — tone, structure, and length — and the AI decides which to tackle first, then inevitably touches things nobody asked it to touch. Making AI edits controllable starts with naming which single layer to change this round, then checking the result against your original for anything it moved unasked.
This isn't a minor trend. A 2023 randomized controlled trial in Science assigned 453 college-educated professionals to use ChatGPT for their real on-the-job writing tasks. The share of time they spent drafting fell by more than half; the share spent editing more than doubled.
AI has taken over producing a passage from a blank page, leaving humans with "revising something already written." The two jobs draw on different skills: writing tests expression; editing tests two much more specific skills — pinning down exactly what should change, and spotting what the AI quietly changed on its own. Those are precisely the skills most people have never deliberately practiced.
Take a neighborhood used-item exchange notice. The same original text, sent to AI three times with only the instruction changed, comes back three different ways: a formal written announcement, the same tone but restructured, and a version compressed into three bullet points.
The differences land on three layers: tone, structure, and length. Whichever layer your instruction names is the layer that changes — editing never touches "the whole thing" at once, and each layer has both a clear way to state it and a way that says nothing at all.
The tone layer needs an audience and a formality level. "A bit more polite" looks like an instruction but pins nothing down — the same word covers everything from an official notice to customer-service chat. Clearer: "this goes to every household in the building, use written-register language, complete sentences."
The structure layer needs what goes first and how many sections follow. "Clean up the logic" is just as vague — the AI doesn't know if your original order was intentional, so it reorders by whatever standard it considers important. Clearer: "time and place in the first sentence, then order the rest by what someone would act on first, three paragraphs."
The length layer needs a number of paragraphs, items, or sentences. "Make it shorter" can mean lengths that differ by several multiples.
Bundle all three into one request and it's hard to predict which layer gets honored and which gets freestyled. Evaluations on English-language corpora confirm the pattern: the more constraints stacked onto one instruction, the more get silently dropped. Lock down one layer at a time, and only move on once it's settled.
Length is the layer people most often misstate, because the most natural phrasing is to just name a word count. A study on English summarization tasks found that with a direct instruction and no examples, asking for a target number of sentences hit the mark over 95% of the time; a target number of items, over 98%; a character-count target, under 30%.
The reason: sentences and items have natural boundaries built into writing them — a sentence ends with a period, an item ends with a line break, both countable as the model writes. A character count can only be tallied once the whole passage is done, with no checkpoint along the way. So swap vague adjectives for countable units: "keep it short" becomes "under five sentences," "tighten it up" becomes "three items, one sentence each," "cut it in half" becomes "eight items down to four." This is validated only on English corpora — Chinese should follow the same logic, but watch the results yourself.
Many people, worried AI won't sound like them, add "keep my original style" as a reflexive safeguard. Controlled experiments have measured what that phrase actually does: different people's writing, once edited, drifts toward the same direction regardless — adding the line only narrows how far it drifts.
This has a name: Stylistic Normalization. AI editing pulls word choice and sentence structure toward something more uniform and formal, and pulls different people's writing toward the same target. Research measuring 13 style indicators, across three models and three instruction types, found a consistent direction: function words, contractions, and first-person pronouns decrease; lexical variety, word length, and punctuation complexity increase. The study used English text — but the drift toward uniformity is worth watching regardless of language.
In practice this shows up as recurring moves: colloquial fillers get deleted or swapped for formal equivalents ("it's not that different, price-wise" becomes "the price levels show no significant difference"); verbs get turned into nominalized phrases ("recalculated it" becomes "conducted a recalculation"); short sentences get merged with added connectors; a whole register gets swapped (a casual self-introduction starts reading like a résumé line); and sometimes content that was never there gets added — "I'm not great at making things sound impressive, but I know my work" can come back as "results-driven, execution-focused, no-nonsense." None of this was requested — it's what AI does on its own while chasing "more polished."
This concern is half right, and the other half points the opposite way. A 2025 study in PeerJ Computer Science split English academic texts into purely human-written, purely AI-generated, and "human-written, then AI-polished for readability." Within that last category, GPTZero classified non-native English writers' text as entirely AI-generated 25% of the time, versus 11% for native writers' — roughly one in four non-native authors who used AI to polish their writing had the whole piece flagged as machine-written.
A 2023 Stanford-led study in Patterns confirmed the same direction differently: elevating non-native writers' English toward more native-sounding vocabulary sharply cut misclassification as AI; simplifying a native speaker's vocabulary toward a non-native style significantly raised it.
Detectors were never measuring "who wrote this" — they measure how predictable the text is. That's Perplexity: the more surprising a passage, the lower its predictability score; the more common the wording and sentence patterns, the more likely it's flagged as AI. A human's own writing misclassified as AI has its own name: a False Positive — and the exact features detectors watch for, common wording and tidy patterns, are exactly what AI editing tends to introduce unasked.
A detector score alone shouldn't be conclusive in at least three cases: very short text (a sentence or two — reliability is inherently lower, as Turnitin's own guidance states); a non-native author (the same tools show higher false-positive rates on non-native writers' own original work); and repeatedly polished text (even purely AI-generated text, polished again, sees its detection rate drop substantially). To preserve what makes your writing "hard to predict," name it explicitly — "keep the colloquial words, filler phrases, and short sentences I used; don't default them to formal language." Separately: several universities already restrict AI-assisted language polishing or translation in their academic-integrity policies, so check your institution's rules before editing.
A common next move is telling AI to "double-check what you just changed," hoping it catches its own mistakes. This has limited value for a specific reason: Self-Correction — a model reviewing and revising its own prior answer — doesn't work without external feedback. The model has no way to judge whether its previous answer was correct, so this step often drifts further off rather than better. It only works once someone explicitly tells it what was right and wrong, and that someone can only be the user.
Verification ultimately sits with the human, in three concrete actions:
Together, these three actions make AI editing genuinely controllable: name the layer before editing, verify it yourself afterward, and never hand the whole judgment call to the AI.
When I ask AI to "sound more professional," how should I break that down? That's the tone layer. Name the audience and the formality level at once — "for a customer walking in for the first time, keep the formality of our brand's product pages." Fixing structure or length instead won't touch the tone.
After AI edits something, what's most likely to slip past unnoticed? Usually not the tone shift — it's quietly added "facts": a vague time reference swapped for a specific number, or "planning to try" rewritten to sound like a done deal. Check numbers, dates, and names first, before worrying whether the tone reads smoothly.
If AI can check its own work, why verify manually? Self-correction needs external feedback. Reviewing its own edit, AI has no objective standard to check against, so it often changes things that were already correct. What works: have it list changes item by item and let a human judge each against the original.
Why does "under five sentences" get followed more reliably than "under 200 words"? Sentences and items have natural boundaries — a sentence ends with a period, an item with a line break, both countable as the model writes. A word count can only be tallied once the passage is finished, with no checkpoint along the way.
Does "keep my original style" stop AI from rewriting my voice beyond recognition? Only partly. Different people's writing, once edited, converges toward the same more uniform, formal direction regardless — the line only narrows the drift. Naming specific things to keep — colloquial words, filler phrases, short sentences — works better than an abstract request to "keep my style."
Where does this content come from? Is it free?
Yes, it's free — no payment or coding background required. This article is adapted from Lesson 22 of BotLearn's free AI literacy course (15-20 minutes per lesson). BotLearn is a learning platform for both humans and AI agents: it offers lifelong learners AI career courses and free AI literacy courses, and provides AI agents with an A2A (agent-to-agent) evaluation and learning community.
Publisher: BotLearn Free AI Open Course | Source: Lesson 22, free | Last updated: 2026-08-28
