Correct, Do Not Rewrite

There is a line running through every AI-assisted draft, and most writers never notice they have crossed it. On one side you ask the model to fix what is wrong: the missing article, the preposition your first language never taught you, the verb that does not agree. On the other side you ask it to make the sentence good — hand it your clumsy paragraph and take back a polished one. The first operation leaves the text yours. The second quietly transfers it to the model. For a non-native English writer the stakes are doubled, because rewriting surrenders two things at once: the authorship you can honestly claim, and the practice that would let you write the sentence yourself next time. Correct, and you keep both. Rewrite, and you rent a result while losing the skill.

This is not a moral scruple dressed up as advice. It is a boundary you can point to, and increasingly one that publishers, detectors, and your own learning curve all draw in the same place.

Two operations that look similar and are not

Start with what actually happens on the page. Correction is local and bounded: you wrote “I have made a research about”, the model returns “I conducted research on”, and you can see exactly what changed and why. The error had a cause — a calque from your first language, a rule you half-knew — and the fix is legible. You could, in principle, produce it yourself once you have seen it. Rewriting is global and opaque: you hand over three sentences of your own reasoning and get back a fluent paragraph whose word choices, rhythm, and even emphasis are the model’s. Nothing shows you what was wrong, because nothing was corrected — the text was replaced.

The distinction matters because the two operations answer different questions. Correction answers “is this right?” Rewriting answers “would a native speaker have written it this way?” — and the honest answer to the second question is almost always no, in the specific sense that a native speaker would have written it their way, not yours. When you accept the rewrite you are not becoming a better writer of your own prose. You are adopting someone else’s, sentence by sentence, until the draft no longer carries your thinking, only your topic.

Where ownership actually breaks

The cleanest account of this boundary comes from research on how writing gets tracked rather than judged. Shenzhe Zhu and colleagues, building a system to certify who wrote what, argue that the final text alone cannot tell you how it was made — and they lay out four writing histories that produce visibly similar documents but carry entirely different claims to authorship (Zhu et al., 2026, p. 2). Two of them sit exactly on our line. In “human draft + AI polish,” the writer produced the substance and the model refined the surface; in “AI draft + human polish,” the model produced the substance and the human tidied it afterward. Zhu’s team notes that a post-hoc reader cannot reliably tell these apart from the document — and that the second case “cannot show that substantive AI generation preceded human editing” (Zhu et al., 2026, p. 2). That is the whole problem in one sentence. Once the model has generated the substance, no amount of your editing on top recovers your authorship of it. The order is what counts, and rewriting inverts it.

Correction keeps you in the “human draft + AI polish” column, where the substantive claim to the work is genuinely yours. Rewriting slides you into “AI draft + human polish” without announcing the move, because the paragraph you sent was yours and the paragraph you got back looks like an improved version of it — not like something the model wrote. It reads as continuity. It is actually a handoff.

The identity cost, measured

For non-native English researchers this is not an abstraction; it is a documented source of strain. Hengzhi Hu and colleagues interviewed 25 non-native English researchers about using ChatGPT in their academic writing and found five distinct “identity configurations,” from reluctant adoption marked by secrecy and moral tension to a settled, reflective integration of the tool (Hu et al., 2025, p. 1). The tension is real precisely because the promise is real: for these writers ChatGPT offers to “mitigate linguistic disadvantage by improving grammatical accuracy, coherence, and fluency,” leveling a playing field long tilted by the global dominance of English (Hu et al., 2025, p. 2). But the same paper records the cost the promise conceals — a growing body of non-native academic work now shows “signs of AI involvement — ranging from stylistic uniformity to the inclusion of fabricated or unverifiable references” (Hu et al., 2025, p. 2). Stylistic uniformity is what rewriting produces at scale: everyone’s paragraphs converge on the same fluent default, and the individual voice that a second-language writer worked years to build dissolves into the model’s register.

There is a telling detail in Hu’s participant profiles. When these researchers self-assessed their skills, they rated themselves stronger at citation and referencing and writing-process efficiency, and weaker at critical evaluation, synthesis, and the organization of arguments (Hu et al., 2025, p. 4). In other words, the parts of writing they most needed to develop were the higher-order, structural parts — exactly the parts a rewrite quietly performs for them and thereby prevents them from practicing. Outsource the sentence and you lose a little polish practice. Outsource the paragraph and you lose the reasoning that the paragraph was supposed to make you do.

That fabricated-references warning deserves its own beat, because it is a concrete hazard of rewriting rather than correcting. When you ask a model to rework a passage that mentions prior work, it can smooth in a citation that looks plausible and does not exist — and it will read as fluently as everything around it. If a rewrite has touched any passage that names sources, run the reference list through a checker that confirms each item is real before you submit; a free bibliography check catches the invented DOI that your eye, trained on the fluent surface, will slide right past.

Correction teaches; rewriting rents

The learning argument is the one that should decide it even if authorship never came up. A correction is a lesson with the answer attached: you see the error, the fix, and — if you pay attention — the rule that connects them. A rewrite is a finished product with the reasoning deleted. You cannot learn from an output that hides its own derivation.

This maps onto a distinction from vocabulary research that we have unpacked before, between the words a learner knows and the control system that retrieves and deploys them under the pressure of real writing. AI help operates on the page, not on that internal system; whether you actually learn from it depends entirely on what you do with the correction, which is the argument at the center of what AI feedback actually improves in L2 writing. A correction you read, understand, and could reproduce feeds the internal system. A rewrite you accept and move past feeds nothing — it delivers a document and leaves you exactly as able, or unable, to write the next one unaided. The friction you remove by rewriting is the friction where learning lives. Every non-native writer who wants to eventually not need the tool has a direct interest in keeping that friction: correct the error, understand the correction, write the next draft with the rule already inside you. This is a habit worth building deliberately, and it belongs to the broader project of writing well as a non-native English writer rather than laundering prose through a model.

Rewriting does not even buy you safety

Writers often reach for the rewrite defensively — the fear is that visibly non-native prose will be flagged as AI, so they let the model “fix” everything to sound native. This backfires twice. First, it is the wrong fix for the stated fear: AI detectors misclassify authentic non-native writing at rates that make them unusable for high-stakes decisions, with controlled evaluations reporting false-positive rates above 60% on non-native academic samples (Doğan & Doğan, 2026). The features that trip these tools — lower syntactic complexity, constrained vocabulary range, repetitive structures — are the ordinary signature of second-language writing, and they “overlap substantially with the textual features that detectors associate with AI generation” (Giray et al., 2025, p. 253). We trace that mechanism in full in why detectors flag non-native English writers: the tool is measuring predictability, not authorship, and second-language prose is predictable by design.

Here is the trap closing. If you respond to that bias by having a model rewrite your draft into fluent native-sounding English, you have now produced text that a process-aware system genuinely cannot distinguish from machine authorship — because in the substance-generating sense it partly is. You have converted a false accusation into a true one to escape it. Giray and colleagues note that detection tools are also “vulnerable to simple evasion strategies, including paraphrasing and translation” (Giray et al., 2026, p. 250), which is the same coin from the other face: the rewrite that hides your voice from a detector is the rewrite that hands your authorship to the model. Correction leaves your genuinely-human draft genuinely human. It is the only move that is honest about both the detector’s failure and your own work.

What the rule looks like in practice

Draw the line at the prompt. Instead of “rewrite this to sound more professional,” ask “mark the grammatical errors and explain each one” — and then make the change yourself. Instead of “improve this paragraph,” ask “which sentences are unclear, and why?” and rework the unclear ones in your own words. Ask for the diagnosis, not the replacement. When the model does propose specific wording, treat it the way you would a colleague’s margin note: adopt the ones you understand and can defend, reject the ones that flatten a meaning you intended. The test is simple — after you accept a change, can you say what was wrong and why the fix is better? If yes, you corrected. If the honest answer is “it just sounds nicer now,” you rewrote, and the sentence is no longer fully yours.

This is also where the emerging rules land, which is worth knowing before a journal or instructor decides it for you. Publishers have begun to formalize exactly this boundary: assistive use that refines your own writing — improving language and readability — is broadly permitted, while generative use that produces substantive content triggers disclosure obligations, and AI cannot be an author because it cannot take responsibility for the work (Elsevier, 2025). Correction sits safely on the permitted side of that line. Rewriting drifts toward the side that, undisclosed, can count as misconduct. The distinction you keep for the sake of your own learning turns out to be the same distinction the institutions are now enforcing.

None of this is an argument against using AI. It is an argument for using it as an instrument of correction rather than substitution — the difference between a tutor who marks your paper and a ghostwriter who replaces it. Keep the model on the surface and the substance stays yours: your authorship holds, your prose keeps your voice, and every session leaves you a slightly better writer instead of a slightly more dependent one. Correct, do not rewrite. The sentence you fix yourself is the only one you get to keep.