Make This Sound Better Tends to Rewrite a Draft Into the Model's House Style

Four things, and none of them is proofreading. Sentence lengths tend to drift toward one band, so the short sentence that carried the point can go. Plain verbs go formal. Hedges and connectives attach themselves to claims the draft had stated flat. And the specific detail only this writer had tends to go as well, because nothing in the prompt told the model that it was the best thing on the page.

What has been measured is the wider loss of variety. A 2025 study across seven datasets and more than 880,000 texts found that language models keep the core content when they polish and rewrite ⁸. What they flatten is the style, and writing-complexity variance falls by 21 to 50 percent ⁸. That study covered polishing and rewriting in general rather than this prompt, and the four changes above are a pattern seen in rewrites rather than a measurement.

The shape is easiest to see on a constructed pair. A student writes:

*The survey only covered 40 households, so the 12 percent figure is shaky.*

A general request to improve that sentence returns the same claim in the register the model treats as better writing:

*It is worth noting that the survey encompassed a relatively limited sample, which may suggest that the 12 percent figure should be interpreted with a degree of caution.*

The sentence went from 13 words to 28. The blunt verb went with it, and four hedges arrived that nobody had written. A reader now has to work out whether the figure is shaky or merely worth some caution.

The table below is drawn from that pair rather than from a measurement.

Not one of those is an error being corrected. They're preferences, and they belong to the model rather than to the student who has to defend the page. A draft can already read this way with no chatbot anywhere near it, which is why an essay can sound like AI.

Because a detector reads the prose and has no view of who typed it, and a rewrite changes the prose. No cited study tested this prompt in ChatGPT, so the nearest evidence is about Grammarly: Bloomberg Businessweek reported in 2024 that using Grammarly to improve an essay or make it sound academic turns work that passed as 100 percent human-written into work read as 100 percent AI-written ¹. Treating a chatbot rewrite the same way is an assumption.

Grammarly's own support pages describe a similar split, though in the score of its own AI detector rather than in the detectors Bloomberg tested. Ordinary corrections should not usually move that score, it says. Its generative features, used to meaningfully rewrite sentences and full paragraphs, will raise it ³. A comma fix and a paragraph rewrite come from the same product and are different kinds of help.

A 2025 study of Arabic writing reported that slight AI polishing of a human-authored article can shift a detector's borderline decision toward AI-generated ⁴. Work published from Stanford in 2023 shows the detectors it tested moving in both directions on word choice alone. ChatGPT rewrote human TOEFL essays with richer vocabulary and the average false positive rate fell from 61.22 percent to 11.77 percent. The same model simplified the word choices in eighth-grade essays from US schools and pushed misclassification the other way, from 5.19 percent to 56.65 percent ². The essays and their arguments stayed the students'. In that study the readings followed word choice rather than authorship, and it says nothing about detectors built since.

What makes the prompt a poor bet is that the effect isn't reliable in either direction. One detector's own figures show how uneven the outcome is, on prompts stronger than this one. Pangram's technical report applied editing prompts meant to substantially modify 14,990 human-written student essays, and its Pangram 4 model read 41.37 percent of the results as fully human, 55.01 percent as mixed and 3.62 percent as AI ⁹. Nature reported the same 41 percent figure from that study ⁵. Whether a plain make-this-sound-better request behaves the same way was not tested there.

So the rewrite might read as machine-written and might not. What the instructor's detector reads is not known until after submission. The version of this accident that starts with a grammar checker has its own page: accused of AI after using Grammarly.

A rewrite hands back prose. A diagnosis hands back a list, and a list is something a student can act on without losing the draft. The request that works names what the model should report rather than what it should produce. Which sentences would a reader have to read twice? Which claims arrive with nothing behind them, and where does a paragraph quietly change subject? The aim is notes, not prose.

The student still writes the sentences. The last row is the distinction Grammarly's own documentation already draws between ordinary corrections, which it says should not usually move its AI score, and generative rewriting, which it says will raise it ³. A chat window runs the same split, though nothing there labels it.

Three habits make the diagnosis more useful. Give the model the assignment prompt and the marking criteria. The notes then answer the question being graded rather than a general idea of good writing. Ask for them a paragraph at a time, because a report on a whole essay tends to come back as a summary, and a summary can't be acted on. And add a line telling the model not to rewrite anything, because a chatbot asked for a diagnosis often appends a suggested rewrite anyway.

A diagnosis also keeps the disclosure short. A student who asked for notes and wrote the sentences has a different thing to declare from one who pasted a paragraph back into the file. That difference fits in a clause. What a course actually permits is the prior question, taken up in whether ChatGPT can edit an essay.

By putting back the two things a rewrite takes out: unevenness and specificity. Read the paragraph aloud and mark where the breath runs out, then break that sentence in half. Then find the one detail nobody else could have written. It might be the figure from the reading, or the example the seminar argued about. Give it the shortest sentence on the page.

The hedges go next. Generally, typically, arguably, to some extent, it is worth noting: cut every one and read what is left. A claim that survives the cut was worth making. A claim that collapses without its hedge was never quite being made.

Openers come next. Write the first two words of every sentence in a column down the margin. A rewritten paragraph tends to show three or four connectives repeating all the way down. The repair is to start on the subject of the claim instead.

One warning belongs here, because the obvious lesson from the numbers above is the wrong one. The same 2024 investigation reported that the students most susceptible to a wrong accusation are likely those who write in a more generic manner, among them neurodivergent students and students who speak English as a second language ¹. Dressing prose up to quiet a detector is no repair either, and the Grammarly result in the same investigation shows where that leads ¹.

Plain and specific is the target, and it's also what a marker is reading for. The longer version of this method, aimed at a draft that did come out of a chatbot, is humanizing AI text by hand.

The syllabus decides first. Where MLA style governs, its August 2025 post asks writers to acknowledge substantive uses of AI ¹⁰; its 2023 post, which MLA has since revised, named editing prose as a functional use to acknowledge in a note, in the text or in another suitable location ⁶. Chicago addresses the reverse case, a writer editing text a tool generated, and asks for a plain phrase such as edited for style and content ⁷.

MLA's August 2025 post puts the reason plainly: acknowledge substantive uses of AI so a reader can tell where the AI ends and the human input begins ¹⁰.

Banning generative tools bans the rewrite whether or not it is declared. A course that allows AI with acknowledgment usually says where the acknowledgment goes. The worked sentence for that note sits at an AI disclosure statement. An integrity meeting costs far more than the clause that would have headed it off, and the clause is easier to write before submission than after a flag.

Where a course allows the tool and a student wants to know which sentences now read as machine-edited, a detector reports that much and nothing else. Human, the detector at human.olive.is, is tuned for college-level academic writing. It reports how much of a text reads as machine-written or machine-edited, sentence by sentence, and separately gives a document verdict of Human, Mixed or AI. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Below 50 words it declines to answer rather than guessing, and between 50 and 149 words a reading is weak because there is so little to read.

A writing tool that reads a draft with the same detector is in private preview: it colors each sentence by what the detector reads it as, leaves quotations, numbers, citations and anything the writer marks as their own untouched, and helps rewrite the rest in the writer's own words, and it never tells a person that their own writing is not theirs. The loop a student can run today works on up to 1,200 words, one paper a day, with no account: check a passage, revise the sentences the detector flags by hand, check again. Human's detector scores the text in memory and de-identifies any personal information for additional training.