Grammarly, Then GPTZero: How ESL Students Get Flagged for Using Grammar Tools

You did the responsible thing. You wrote your essay in English, your second or third language, then ran it through Grammarly to clean up the articles and verb tenses you weren’t sure about. Two days later your professor emails you: GPTZero says your paper is 92% likely AI. You didn’t use AI. So what happened?

This is one of the strangest traps in academic writing right now. The tool you used to look *more* careful is part of why a detector decided you looked *less* human.

Key takeaways

• Grammar tools smooth out the exact irregularities that AI detectors read as “human,” so heavy editing can nudge an honest draft toward an AI flag.
• Detectors score statistical patterns, not honesty. Accepting every suggestion makes your writing more predictable, and predictable reads as machine-like.
• ESL writers get hit harder because they already favor safe, common phrasing, and grammar tools push that even further toward the regularized output detectors associate with AI.
• A flag is a probability, not proof. Keep your drafts and version history so you can show your process.
• Use grammar tools for real errors, but protect your own sentence variety and word choices.

What AI detectors actually measure

Here’s the thing most people get wrong: detectors like GPTZero don’t have a secret list of “AI sentences.” They can’t see your intent, and they don’t know whether you cheated. They measure how *predictable* your text is.

The main signal is something called perplexity. Low perplexity means each word is roughly what a language model would have guessed next: smooth, expected, statistically boring. AI text tends to have low perplexity because that’s literally how the models are built. They pick high-probability words. The second signal is burstiness, the variation in sentence length and rhythm. Humans write in jagged bursts. A long, winding sentence, then a short one. AI tends to hum along at an even, mid-length clip.

The short version: detectors reward irregularity and punish smoothness. Now look at what a grammar tool is *designed* to do.

Grammar tools optimize for the wrong thing (for detection)

Grammarly, and tools like it, exist to make your writing cleaner, clearer, and more conventional. That’s a good goal for communication. It’s a terrible goal for not-looking-like-AI.

When you accept a suggestion, you’re usually trading an unusual choice for a common one. A slightly odd phrasing becomes a standard phrasing. A long, slightly rambling sentence gets tightened. A word the tool thinks is “wordy” gets cut. Each individual edit is reasonable. But stack fifty of them on top of a 700-word essay and you’ve quietly lowered the perplexity and flattened the burstiness, which are the two things a detector is watching.

Grammarly isn’t generating text, so it doesn’t “tag” your work as AI. The damage is indirect. You’re hand-feeding your draft through an optimizer that pushes it toward exactly the statistical profile detectors learned to associate with machines. The tool that’s supposed to prove your diligence ends up sanding off your fingerprints.

A quick before-and-after

Say an ESL student writes this line in a first draft:

> “The factory it was closing down, and many of the workers they did not know where to go next, which made the whole town feel kind of frozen.”

Messy? A little. It has a run-on quality, a doubled subject, and an informal “kind of.” But it’s *bursty*, and the word choices are slightly unexpected. A detector might well read it as human.

After accepting every Grammarly fix, it becomes:

> “The factory was closing down, and many workers did not know where to go next. This uncertainty left the entire town feeling frozen.”

Grammatically? Much better. Cleaner, correct, more “academic.” But it’s now two even, well-formed sentences full of high-frequency phrasing: “left the entire town feeling,” “did not know where to go.” It reads smoother, and smoother is precisely what tilts a perplexity score toward “AI.”

The student improved the grammar and raised the AI score at the same time. Nobody told them those two things could pull in opposite directions.

Why this hits ESL writers harder

Native speakers and non-native speakers don’t get flagged at the same rate, and there’s research on it. A 2023 study published in the journal *Patterns*, led by Weixin Liang and James Zou at Stanford, ran seven popular detectors against essays written by non-native English speakers. More than half of the TOEFL essays, written by real students, got labeled “AI-generated,” an average false-positive rate of around 61%. For essays by US eighth-graders, the same detectors cleared over 90% as human. The reason maps cleanly onto perplexity, and the authors say as much: simpler, more predictable word choices are exactly what these tools penalize.

Think about why that happens. If English isn’t your first language, you’ve probably been trained, correctly, to use clear, standard constructions and reliable vocabulary. You reach for the word you’re *sure* is right, not the quirky one. You repeat sentence patterns you know are grammatical. That caution is smart. It’s also low-perplexity by nature. Your writing already sits closer to the “predictable” end of the scale before you touch any tool.

Then you add a grammar checker that nudges you even further toward standard, expected, conventional phrasing. A native speaker who writes with slang, sentence fragments, and weird rhythm has perplexity to spare. The careful ESL writer who runs everything through Grammarly has almost none left. Same tool, very different risk.

It’s not fair, and it’s not your writing’s fault. The detector is reading carefulness as a confession. We’ve written more about why GPTZero sometimes flags human text if you want the mechanics.

What to actually do about it

You don’t have to choose between fixing your grammar and avoiding false flags. You just have to use the tools with your eyes open.

Accept corrections, reject rewrites. When Grammarly flags a real error, a missing article, a wrong tense, a true typo, take the fix. When it suggests a “clearer” or “more concise” rewrite that changes your phrasing, pause. If your original was correct, keep it. Your slightly less polished version is often the more human one.

Protect your sentence variety. After editing, reread and check the rhythm. If every sentence is suddenly the same length, break one up or combine two. Put a five-word sentence next to a thirty-word one. Burstiness is your friend.

Keep your own words. That unusual verb you chose, the specific concrete detail, the example only you would think of, those are exactly what reads as human. Don’t let a tool sand them off in the name of “clarity.”

Save your trail. Write in something with version history, like Google Docs. If you’re ever questioned, a revision history showing the document grow from a rough draft to a finished essay is far stronger proof of authorship than a clean final file.

Read the score for what it is. A detector outputs a probability, not a verdict. An “AI likely” result is a statistical guess that can be, and often is, wrong. OpenAI quietly retired its own AI classifier in July 2023, citing a low rate of accuracy. If a free tool from the company that *makes* GPT couldn’t reliably detect AI text, your professor’s third-party checker isn’t infallible either.

If a flag does land on you, knowing your rights matters as much as fixing the prose. We put together a guide on what to do when detectors get it wrong and how students can defend themselves, and there’s more on detection and humanizing if you want to dig into patterns by writing style and language background. Curious what the tool costs? Here’s how PaperBleach is priced.

Frequently asked questions

Does using Grammarly cause AI detection flags?

Grammarly isn’t an AI writing generator, so it doesn’t mark your text as AI directly. But its edits can raise your score. Detectors measure how predictable your writing is, and when you accept every suggestion, your sentences get smoother, more uniform, and full of common phrasing, which is the same statistical signature these tools associate with machine output. It’s not that Grammarly flags you. It’s that over-edited text reads as low-perplexity, and that’s what gets flagged.

Why are ESL students flagged more often than native speakers?

Non-native writers tend to stick to safe, high-frequency words and standard structures, both because that’s what they were taught and to avoid mistakes. Detectors treat that predictability as a sign of machine generation. The 2023 *Patterns* study led by Weixin Liang and James Zou found seven detectors misclassified more than half of non-native TOEFL essays as AI (about a 61% false-positive rate) while clearing over 90% of US eighth-grade essays. Add a grammar tool that regularizes the text further, and the risk compounds.

Is an AI detection score proof that I cheated?

No. Detectors output a probability, not a verdict. “85% AI” means the text’s statistics resemble what the model has seen in AI writing. It is not evidence you used AI. These tools produce false positives, which is part of why OpenAI retired its own classifier in July 2023 for low accuracy. If you wrote the work, the score is simply wrong, and you can push back.

How can I prove I wrote my essay if Grammarly got me flagged?

Keep your trail. Use a tool with version history, save earlier drafts, and hold onto your outline, notes, and sources. A document you can show growing over time, from a messy first draft through edits and revisions, is far stronger evidence than a clean final file. Being able to explain your argument in your own words in a short conversation helps too.

Should ESL students stop using grammar tools entirely?

No. They’re genuinely useful for catching real errors: agreement, articles, tense. The fix is how you use them. Accept corrections for actual mistakes, but don’t accept every stylistic rewrite. Keep your own sentence rhythm, mix short and long sentences, and hold onto word choices that are correct even when the tool prefers something “smoother.”

The bottom line

There’s a bitter irony here. The more carefully you try to write, the more some detectors suspect you. That’s a flaw in the detectors, not in you. Fix your real grammar mistakes, keep your own voice and rhythm, save your drafts, and remember that a score is a guess, not a judgment of your honesty. If you want to see exactly which sentences are tripping a detector before you turn anything in, run your draft through a humanizer and edit with the patterns in plain sight.