Why AI Detectors Flag Non-Native English Writing

The writing that gets you flagged isn't sloppy. It's careful. Simple sentences, common words, no risk-taking, exactly what a cautious non-native speaker produces on purpose, and exactly what an AI detector has learned to treat as a warning sign.

That's not a guess. It's a published, peer-reviewed result, and the gap is bigger than most people assume.

The study behind this problem

Back in 2023, a Stanford team led by James Zou tried something simple. Take 91 real TOEFL essays, essays written by non-native English speakers for a standardized test, and run them through seven commercial AI detectors. The results ran in the journal Patterns, and they were stark. On average, the detectors flagged 61.3% of the TOEFL essays as machine-written. Native-English writing samples got flagged 5.1% of the time. Eighteen of the ninety-one essays, that's 19%, got flagged by all seven detectors at once, unanimous agreement. Eighty-nine of the ninety-one, 97%, got flagged by at least one.

Those are essays written entirely by humans, before ChatGPT was in wide use, submitted to prove English proficiency for university admission. The detectors weren't confused by AI. They were confused by a writing style.

Why "simple" reads as "synthetic"

The mechanism has a name: perplexity. It measures how surprising each word choice is, given the words around it. Language models generate the statistically expected word by design, so their output has low perplexity. Detectors learned to treat low perplexity as the signature of AI text.

Non-native writers do the same thing for a completely different reason though. Call it caution. When you're not fully confident in your English, you reach for the word you know is correct instead of the more distinctive one you might get wrong. "Big" instead of "substantial." Shorter, safer sentences instead of the kind with three embedded clauses. Playing it safe like that produces exactly the low-perplexity signature detectors were built to catch.

The tragedy is that the two writers, a language model minimizing surprise and a nervous student minimizing risk, end up looking statistically similar to a machine that only measures the shape of the sentence, never the reason behind it.

Here's what that looks like on the page. A cautious sentence might read: "The economy is very big and it is growing fast." Safe, correct, and low-perplexity. A sentence with more native-level risk-taking might read: "The economy has ballooned, expanding at a pace few forecasters predicted." Same idea. The second version takes chances with word choice a careful non-native writer is trained to avoid, and it's exactly the kind of unpredictability that keeps a detector from flagging it.

Notice what that means in practice: two students can express an identical thought, and only the one who plays it safe with language gets treated as suspicious. That isn't a writing-quality problem. It's a bias built into how the detection model was trained.

The fix Stanford already tested

The same study tried an intervention: they had ChatGPT rewrite the TOEFL essays with instructions to enhance the vocabulary "to sound more like that of a native speaker." Richer word choices raised the perplexity score, and the misclassification rate fell from 61.3% to 11.6%.

I tell students this because it's a tested result, not a hunch. Deliberately vary your vocabulary. Mix up your sentence length. Stop defaulting to the safest phrasing every time you sit down to write. None of that erases the bias built into these tools. It just cuts down, measurably, how often you get caught in it.

Where policy stands in 2026

The response from universities has been uneven but moving in one direction. Several, including Vanderbilt, MIT, and Yale, disabled AI detection entirely rather than adjudicate a tool with a documented bias against a protected class of student. Elsewhere, the more common shift has been away from blanket "AI is banned" rules toward disclosure-based policy: undisclosed AI use is the violation, not AI-adjacent writing style.

Neither shift undoes a flag that already happened to you. If your school still runs detection and you get one, the score itself isn't proof of anything, a point worth understanding fully since Turnitin's own documentation says the same thing about what that percentage does and doesn't mean.

If you're an international student, a false flag hits harder than it would for a classmate born here. A finding of academic dishonesty can shake a student visa's standing. It can cost a scholarship tied to your GPA. It can undo a conditional admission that assumed a clean record. None of that means panic. It means treating the documentation step below as something you actually do, not something you mean to get around to.

What "vary your vocabulary" looks like in practice

This advice is easy to give and hard to act on if you're not sure what to change. A few concrete moves that raise perplexity without changing what you're trying to say:

Trade one safe word per paragraph for a riskier one you already know is correct. Not a thesaurus dive, just the word you'd normally avoid because a simpler synonym feels safer. "Increased" becomes "climbed." "Difficult" becomes "grueling," when the sentence earns it.
Let one sentence run long. If every sentence in a paragraph is eight to twelve words, that uniformity itself reads as cautious. One longer sentence with a subordinate clause breaks the pattern.
Keep a phrase bank from your reading. When you read something in English, native or not, that phrases an idea in a way you wouldn't have thought to, write it down. Reuse it later where it fits. This is exactly how vocabulary confidence gets built, one borrowed phrase at a time.
Read your draft aloud. If every sentence sounds like it could have come from a textbook exercise, that flatness is the signal a detector is picking up on too.

What to actually do about it

A few things that help, in order of effort:

Vary your vocabulary on purpose. Not to sound fancier, to sound like a writer who takes risks with word choice instead of always retreating to the safest option.
Mix your sentence lengths. A paragraph of uniform, short, safe sentences is the exact pattern both AI output and cautious ESL writing share.
Check your own draft before you submit it. A free AI detector won't predict your school's exact tool, but it tells you whether your writing currently sits in flagged territory, while you still have time to revise.
Keep your process. Notes, outlines, drafts with timestamps. If a flag does happen despite everything, this is what actually resolves it, not arguing about vocabulary.

If a flag has already landed and you're facing a conversation about it, there's a specific process for proving you wrote something yourself, built around exactly this kind of evidence.

One more thing worth saying directly to instructors reading this: if a student's writing style is consistent across a semester of graded work, that consistency is itself evidence. Comparing a flagged paper against work the same student turned in earlier is a five-minute check that catches most false positives before a formal process ever starts. A one-time detector score should never outweigh a semester of a student's own writing history.

Know the number before someone else uses it against you

This isn't a fringe complaint. It's a Stanford study, peer-reviewed and published, showing a twelve-fold gap in false-flagging between native and non-native English writing. If you're an international or ESL student, that bias applies to you specifically. Knowing the exact numbers, 61.3% versus 5.1%, is worth more walking into a meeting than any amount of nervous apologizing.

Write with the vocabulary you actually have. Vary it where you can. And if the flag comes anyway, you now know it was never really about you.