Does Proofreading Your Essay Make It Look AI-Written?
September 8, 2026 · PT Technologies · 6 min read
You wrote the essay yourself. Then you ran it through a grammar checker, because the commas were a mess and one sentence had been bothering you for two days. Now you are wondering whether that innocuous step has made your own work look machine-written.
It is a reasonable worry, and the honest answer is yes, it has some effect — it is measurable, we have measured it on our own detector, and the numbers are worse than most people expect.
What we measured
The test that matters here is a paired one: take the same essays, score them, proofread them, and score them again. Anything else confounds the effect with differences between texts.
We built exactly that. 456 learner essays, each scored before and after light proofreading — light meaning the edited version stays 75 to 99.5% word-identical to its source, with a median similarity of 0.93. That threshold is doing real work: anything below it is not proofreading, it is rewriting, and we excluded those runs rather than let them muddy the result.
Under our v2.9 model, scoring those identical essays before and after:
| Before proofreading | After proofreading | |
|---|---|---|
| Essays flagged | 1.32% (6 of 456) | 3.73% (17 of 456) |
Eleven essays crossed the line purely because they had been proofread. Nothing about the ideas, the argument or the authorship changed. Running a proofreader over your essay nearly tripled your chance of being falsely flagged.
That is the thing worth sitting with. The step everyone tells students to take — check your grammar before you submit — was actively working against them.
Why it happens
Detectors do not detect AI. They detect statistical properties that correlate with generated text: how predictable each word is given the ones before it, how much sentence length varies, how uniform the rhythm is.
Proofreading moves a text along exactly those axes. It regularises punctuation. It replaces the odd construction with the conventional one. It smooths the sentence that ran long and fixed the one that ran short. Every one of those edits is an improvement to a human reader and a step toward the statistical profile of machine text.
The mechanism is not a bug in any particular detector. It follows from what they measure. Any tool built on predictability and rhythm will show some version of this effect, which is why the finding generalises even though the numbers below are ours alone.
What we changed
The gap was in the training data, and it was a specific one. Our AI class already contained AI-written-then-edited text. Nothing anywhere contained the opposite and far more common case: a human draft run through a proofreader. Nothing in training had ever said those are still human.
So we built that class — 1,939 human texts lightly edited under six real proofreading instructions, gated on the same similarity threshold. 79% of it is learner English, deliberately: the intersection of "most likely to use a grammar checker" and "most likely to be falsely accused" is precisely where the risk concentrates, and building the class from the average case would have missed it.
On the same 456 paired essays, after that change:
| v2.9 | v2.10 | |
|---|---|---|
| Flagged before proofreading | 1.32% | 0.66% |
| Flagged after proofreading | 3.73% | 1.75% |
| Newly flagged because proofread | 12 | 5 |
| Score penalty for proofreading (p90) | +0.257 | +0.048 |
The penalty for proofreading is now about a third of what it was, and the score movement itself is five times smaller at the 90th percentile.
What that does and does not mean
It is still not zero. 1.75% of 456 essays is eight, and five of those were flagged only because the author did something every writing guide recommends. If you are one of the eight, the improvement is cold comfort. We publish the number rather than the improvement because the number is what affects you.
These are our figures, for our detector. We cannot measure Turnitin's, GPTZero's or anyone else's from the outside — they do not expose the paired evaluation you would need. The mechanism is general, so we would expect the effect to exist everywhere; the size of it we genuinely do not know, and nobody should claim otherwise about a tool they cannot test.
A low rate across a corpus is still a lot of people. This is the same arithmetic we set out in a 1% false positive rate is not small: a per-essay rate that sounds like a rounding error becomes a headcount once you multiply by a cohort.
What to do about it
Keep your drafts. Version history in Google Docs or Word, or dated files. A visible progression from notes to finished essay is stronger evidence of authorship than any score is evidence against it, and it is what actually persuades people. This is the single highest-value habit in this whole area and it costs nothing.
Edit locally. Fix the comma splice; do not hand over the paragraph. Sentence-level corrections move the statistics far less than wholesale rewriting, and they keep the writing sounding like you — which is the point of writing it.
Do not stop proofreading. This is worth saying plainly, because the wrong lesson to take from the table above is "submit it messy". Unclear writing costs you marks with certainty; a false flag is a small probability you can mitigate other ways. Proofread, keep the drafts, and if it comes up you have the evidence.
Look at what got flagged, not just the number. If you want to see how your writing reads to a detector, our AI Detector shows the score with its margin of error and the specific sentences driving it. A high score concentrated in your methods section or your literature summary usually means those passages are formulaic — which is what those passages are supposed to be, and is not the same finding as "this was generated".
The broader point
The reason we ran this evaluation at all is that the failure mode had never been measured, and unmeasured failure modes do not get fixed. It is easy for a detector vendor to quote a headline accuracy figure and never test the one situation that produces the most unfair outcomes.
If you were flagged and you wrote every word, flagged as AI when you wrote it walks through what to do next. And if you are wondering where the line sits on using these tools at all, we took that question on separately in is using AI to fix your grammar cheating?.