Rewrite the Sentences That Read as AI, Not the Whole Essay

Because the signal the studies below measured is style, not honesty, and every source that explains it points at word choice. Bloomberg Businessweek reported in 2024 that detectors typically look at how complex the words are, and it quoted the Stanford study's senior author: generic, formulaic word choices stand a higher chance of being flagged ¹. Deadline prose can be exactly that.

The 2023 Stanford study's own reading of its result is that detectors may penalize writers with a narrower range of expression ². The experiment behind that sentence is worth knowing by number. Seven detectors available in early 2023 read 91 TOEFL essays written by people, and more than half came back as machine-written, an average false positive rate of 61.22 percent ². At least one of the seven flagged 89 of the 91 ². No machine had touched those essays. No source cited here reruns that study on current detectors, so its rates describe the tools of early 2023.

Bloomberg Businessweek arrived at the same place from the other direction in 2024. The students most exposed to a wrong accusation, it reported, are likely the ones who write in a generic way, according to the students, academics and developers it spoke to: neurodivergent students, students writing in a second language, or anyone who learned a straightforward vocabulary ¹.

One vendor names the cases it recommends against screening at all, in its own documentation. Pangram Labs, which sells a detector, puts outlines, single-sentence answers and template writing on that list, because it is most confident in its accuracy when text is long, in complete sentences and needs more original input from the writer ⁵.

A deadline produces plain material. An essay drafted the night before, from a plan, in the register a rubric asks for, comes out even, hedged and plainly worded, and plain wording is what got flagged in the Stanford set ². That set was 91 TOEFL essays by second-language writers. Whether a fluent writer's hurried draft is read the same way is not measured in any source here, so the step from one to the other is an assumption.

Where the writer's first language is not English the effect is the one that has been measured, and it has a page of its own: a non-native speaker flagged by a detector.

Locating is most of the job, because a rewrite aimed at nothing in particular smooths the essay and changes the wrong lines. Two routes find sentences to reopen, and they answer different questions. A detector that reports sentence by sentence marks what a model reads as machine-written. A pencil finds the stretch a reader experiences as flat: write each sentence's word count in the margin for one page, and the flat stretch shows itself.

Four habits make a page read flat to a person. Every one of them is countable without software.

A detector adds the one list the pencil cannot make: the sentences a model reads as machine-written. Nothing cited here compares that list with the one a careful writer would guess, and no source in this article says which features a detector weighs beyond word choice, so the pencil list is a reader's list, not a detector's. Read the marks on either list as places to look.

Turnitin published the arithmetic that makes that caution necessary. It said in June 2023 that its sentence-level false positive rate was around 4 percent, roughly one highlighted sentence in every 25 being human-written, and it told instructors to use the information to start a conversation, not to draw a conclusion ³.

Position can change a reading as well. Max Spero of Pangram Labs told Nature in 2026 that, in its tool, a paragraph judged human on its own can switch to an AI reading once it sits inside a longer segment ⁴. Pulling one sentence out and scoring it alone answers a question nobody asked.

Close the draft. Write the marked claim again from what the student actually knows, then reopen the file and keep only the facts and citations the old sentence was carrying. The rebuilt sentence reads as one person's, which is a change a synonym swap inside the old one does not make.

Swapping words inside the sentence is the re-wording that moved readings in both directions in the Stanford experiments ², so it chases the number. Rebuilding from notes is worth the extra step for a different reason: the result is the student's own sentence, whatever a reading says about it.

Two things change when a sentence is rebuilt from memory. It comes back shorter, and its length stops matching its neighbors, because memory does not produce even prose. And the hedge bolted to the front of it falls off, because a claim recalled is a claim the writer is willing to make. Hedging feels careful. At volume it reads as evasive.

What has to survive the pass is narrow. Every figure. Every citation marker and every quoted span. The claim itself. A rewrite that quietly loses a citation has damaged the essay.

The test at the end of each rebuilt paragraph is whether the student could say it out loud in an office hour, without the page in front of them. A paragraph that fails that test is not ready, whatever any instrument reports about it.

The same pass run over a whole machine-written draft instead of a few marked sentences is a different job with different consequences: humanizing AI text by hand.

A polish pass, in the one test this article cites. Bloomberg Businessweek ran a human-written essay through Grammarly's rewriting features in 2024 and watched it go from reading as 100 percent human-written to 100 percent machine-written, while the plain spelling and grammar fixes barely moved it ¹. A model asked to improve a paragraph rewrites it in the model's own register. Where the reading lands afterwards is out of the writer's hands.

The Stanford result runs in both directions. Having ChatGPT enrich the vocabulary of human TOEFL essays cut the average false positive rate from 61.22 percent to 11.77 percent ². Having ChatGPT simplify the word choices in essays by eighth graders in the United States pushed misclassification the other way, from 5.19 percent to 56.65 percent ². In both experiments the re-wording was done by ChatGPT ².

Grammarly's polish raised the reading in Bloomberg's test ¹, and ChatGPT's vocabulary enrichment lowered it in Stanford's ². A number that moves that far on a re-wording is tracking word choice at least as much as who wrote the essay, and editing toward it teaches a student to dress prose rather than write it. The chatbot version of the request tends to arrive the same afternoon, and it has its own page: make this sound better.

One more reason to stop. The seven detectors in the Stanford set disagreed with each other: at least one flagged 89 of 91 essays while the average was 61.22 percent ². A free checker also differs from whatever instrument a course actually runs, so a draft tuned to one reading is tuned to a number nobody will grade with. That cycle is the detector and humanizer loop.

Read the revised draft through the same instrument and treat the second reading as a record of what changed, never a score to chase. A draft over the paste limit goes in as the same section both times. Short extracts are the case to avoid. A detector handed 40 words of a paragraph has almost nothing to read, and a weak reading on a fragment gets quoted later as though it were a strong one.

The instrument this article assumes for that second pass is Human. At human.olive.is a student can paste up to 1,200 words and get a reading, one paper a day, with no account, and it reports sentence by sentence how much of the text reads as machine-written or machine-edited, with 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. Human's detector scores the text in memory and de-identifies any personal information for additional training. Below 50 words it declines to answer, and between 50 and 149 words a reading is weak because there is so little to read. Pangram Labs also recommends against screening very short responses ⁵.

A writing tool built on the same detector is in private preview. It colors each sentence by what the detector reads it as, keeps quotations, numbers, citations and anything the writer marks as their own untouched, and helps rewrite the rest in the writer's own words; it never tells a person that their own writing is not theirs. What a student can run today is the loop this section describes: check the draft, revise the marked sentences from notes, check again.

Keep the evidence the revision produced. It outlasts any score. William Quarterman, a UC Davis senior, was referred for academic dishonesty after his professor used AI detection software, including GPTZero, in a case USA TODAY reported in 2023 ⁷. The edit history of his Google document cleared him: it carried time-stamped proof of the paper being built over two or three hours ⁶. A revision written in a document with version history leaves that trail without any extra work: version history as proof.

Where a tool did touch the prose, the course policy decides what has to be said about it. One sentence in the right place costs less than an integrity meeting, and the wording has a page of its own: an AI disclosure statement.