An AI humanizer for ESL writers has to start somewhere uncomfortable. If you write English as a second language and a detector flagged your own work, you are not imagining a pattern. One study measured an average false positive rate of 61.3% across seven detectors on TOEFL essays by non-native writers, their measurement rather than ours (Liang et al., 2023). Native-written control essays in the same study were barely flagged at all.
That shapes what this page recommends, and it is not what you would expect from a company selling a rewriter. If you wrote the text yourself, rewriting it is usually the wrong first move. A rewrite replaces the evidence that you wrote it. Keep reading for what to do instead, and for the narrower case where a humanizer genuinely helps an ESL writer.
It is not about mistakes. It is about predictability, and that is the uncomfortable part.
Detectors read two things: how predictable your word choices are, and how much that predictability varies between sentences. Low variation plus high predictability reads as machine.
Second-language writing often scores exactly that way, for reasons that have nothing to do with quality. A learner draws on a smaller, more carefully chosen vocabulary. Sentence structures stay closer to the patterns that were taught, because those patterns are reliable. Idiom is used sparingly, since getting idiom wrong is costly. Every one of those is good discipline. All of them reduce the statistical surprise in your text.
So the signal a detector reads isn't "this person used AI". It is "this text is less surprising than a native-speaker average". Careful ESL prose and model output can look alike on that one axis while being nothing alike in origin.
This is a documented weakness, not a theory. GPTZero names second-language English among its own weak cases. We publish the same limitation on our detection limitations page, because a score you can't interrogate is worth very little.
The instinct is to run it through a humanizer. That instinct makes your position worse, and here is the specific reason.
A rewrite replaces the thing you are defending. If your own sentences are gone, the draft history, the phrasing you can explain, and the notes that match your text are gone with them. You are left asking someone to trust a version you didn't write either.
What actually helps is process evidence. Version history in the editor you used. Earlier drafts. Your notes and outline. The ability to talk through your argument and say why you chose a structure. That evidence is far stronger than any score, and how to prove you didn't use AI sets out what to gather.
Then cite the research. The 61.3% figure above is the single most useful thing an ESL writer can bring to a conversation about a false flag, because it moves the discussion from your character to a known property of the tool. Our write-up on detector bias against non-native writers is written to be forwarded to an instructor.
One honest caveat. None of this guarantees an outcome, and we're not going to pretend otherwise. It gives you something concrete to show instead of a denial.
Three situations, all of them before submission rather than after an accusation.
You used a model to help draft, and you want the result to read like you. This is the legitimate and common case. The output is fluent and characterless; your job is to put yourself back into it. A rewrite varies the rhythm, and then you edit for the detail only you know.
You are writing in English for the first time in a professional register and the draft reads stiff. A rewrite shows you an alternative phrasing of your own sentence, which is a genuinely useful way to learn register. Read the two side by side and pick deliberately.
You want to know how your writing scores before you submit it, not after someone else runs it. Scoring is free and separate from the humanizer, and for an ESL writer it's the more valuable half.
In all three the humanizer is an editing aid you drive, not a verdict you accept. That is the only framing under which a humanizer is worth an ESL writer's time.
There are three: Light, Balanced and Maximum. For ESL writers, start at Light and usually stay there. Light works on sentence rhythm and cuts connective filler, which is where the detector signal actually sits. Balanced edits phrasing more broadly. Maximum replaces most of the wording and is a Pro mode; the free tier covers Light and Balanced.
The reason to stay light is specific to second-language writing. A heavier rewrite reaches for vocabulary, and a synonym that fits the sentence may not fit the field or the register you were aiming for. You can end up with text that is less yours and no more defensible. Humanizing without changing meaning covers the check in detail.
We won't tell you a rewrite makes your work pass a detector. We can't see another tool's verdict, nobody can, and any product claiming a guarantee against a system it doesn't control is selling you the sentence rather than the outcome.
What we do is score your text before the rewrite and after it with our own detector, label the reading as ours, and highlight which words changed. That tells you whether the statistical properties this whole family of detectors reads have moved. It isn't a prediction of what your institution's tool will say.
We also won't pretend the underlying problem is yours to fix. Detectors over-flagging second-language English is a flaw in the detectors. A rewriting tool is a workaround, and workarounds are worth having, but it would be dishonest to sell one as a solution to somebody else's measurement error.
Most likely because careful second-language English is statistically less surprising than a native-speaker average, and that is the property detectors measure. A smaller deliberate vocabulary, reliable sentence patterns and sparing idiom all reduce variation, and low variation reads as machine-written. Liang et al. (2023) measured an average false positive rate of 61.3% across seven detectors on TOEFL essays by non-native writers while barely flagging native-written controls. It is a known weakness of the tools, not evidence about you.
Usually not. A rewrite replaces the sentences you are trying to defend and destroys the draft history that would have supported you. Gather process evidence instead: version history, earlier drafts, notes, and the ability to explain your argument. Cite the research on detector bias so the conversation moves from your character to a known property of the tool. Use a rewriter on the next draft while you are writing it, not on a draft already under question.
Light, in almost every case. Light works on sentence rhythm and removes connective filler, which is where the detector signal actually sits, and it leaves your vocabulary alone. Balanced edits phrasing more broadly and suits general prose. Maximum replaces most of the wording and is a Pro mode; the free tier covers Light and Balanced. Heavier rewriting is riskier for second-language writers because a substituted synonym can miss the register or the technical sense you intended.
Our detector is built and measured for English, so the before and after scores are only meaningful on English text. The rewriter will accept other languages but we do not publish quality or accuracy claims for them, and you should not rely on the score. If you draft in another language and translate into English, score the English version, since that is what will be read and what will be checked.
It depends entirely on your institution and what you are doing. Editing your own draft for clarity and rhythm is ordinary writing practice and most policies treat it that way. Generating text with a model and presenting it as unaided work is a different act, and a rewrite does not change what it is. Check your institution policy, disclose AI assistance where the policy asks for it, and remember that a tool cannot give you permission your institution has not.
We do not publish a single number for this, because an honest one would need to name the detector, the version, the text type and the writer population, and it would go stale quickly. The best-documented figure in the published record is the 61.3% average across seven detectors on TOEFL essays from Liang et al. (2023). For our own detector we publish what we measure and what we do not on the detection limitations page rather than quoting a convenient figure.
Without an account, 3 rewrites a day at up to 300 words each, which is enough to see how your own writing behaves before deciding anything. A free account gives 260 words per rewrite in Light and Balanced modes. Scoring your text is the part most ESL writers actually want, and you can do that before you rewrite anything.
The published evidence, written to be forwarded to an instructor.
Read the evidence →The process evidence that is worth more than any score.
See the evidence list →What drifts first in a rewrite, and how to check your argument survived.
Read the method →What our detector cannot tell you, published rather than buried.
See the limits →The full tool, with the before and after scored by our own detector.
See the tool →What the number measures, and what a before and after reading can tell you.
Understand the score →Read your own writing the way a detector reads it, before anyone else does. 3 scores a day, 300 words each, no account. If you wrote it yourself, scoring is the useful half and rewriting is the part to think twice about.