Do AI Detectors Punish ESL Students? What the False-Positive Data Actually Shows

If you teach ESL or EFL writing, you've probably already had the conversation: a student's essay gets flagged by an AI detector, the student swears they wrote every word, and you're left trying to judge a percentage score you don't fully trust. You're right to be skeptical. A growing body of testing shows that popular AI detectors misfire on non-native English writing at rates far higher than they do on native writing — and the reasons come down to how these tools actually work, not how honest your students are.

What the false-positive numbers actually show

Independent testing of AI-detection tools has found meaningfully different false-positive rates depending on which tool you use and whose writing you feed it. Figures reported in that testing put GPTZero's false-positive rate on non-native English writing at around 38%, Turnitin's at around 18%, and Copyleaks' at around 13%. Native-English writing samples in the same testing came back clean far more consistently.

Those aren't rounding errors. A false-positive rate anywhere near 38% means that for every three students whose writing gets flagged, more than one of them may have written it entirely themselves. If you're using a detector score as anything close to a verdict, you are very likely accusing innocent students — and the students most at risk are exactly the ones you're there to support.

Why non-native syntax trips these detectors

AI detectors work by estimating how "predictable" a piece of text is — how closely word choices and sentence structures match patterns the model has seen before. Native fluent writing tends to vary in ways detectors have learned to expect. Non-native writing often doesn't, for reasons that have nothing to do with AI.

Learners at the same CEFR level frequently produce similar sentence structures because they were taught the same structures. A B1 class working from the same textbook will naturally converge on similar transitions, similar hedging phrases, similar ways of opening a paragraph. That's not a red flag — that's what learning a language from a shared curriculum looks like. But to a detector trained mostly on native writing, that convergence can look statistically "unnatural" in the same direction AI-generated text does.

Simpler vocabulary and more restricted sentence variety — both completely normal at earlier CEFR levels — push the same signals. The detector isn't reading for authenticity. It's reading for statistical smoothness, and non-native writing is often smoother in the wrong way for entirely legitimate reasons.

What this means for your grading policy

The practical takeaway isn't "don't worry about AI-written work" — plenty of students do use generative AI to write assignments, and that's a real concern. The takeaway is that a single detector score is not reliable enough to act on by itself, especially with ESL writers. If your current policy treats a detector percentage as the deciding factor, it's worth revisiting before it costs a student their grade — or your class its trust in you.

A more defensible approach separates suspicion from judgment. Detector output can be one input among several. It should never be the entire case.

Process-based review beats a single AI-detector verdict

The more reliable path is to look at how a piece of writing was produced, not just guess at the finished text. That usually means paying attention to things a detector can't see: whether a student's drafts show visible revision over time, whether their writing pace and rhythm match how they normally work, and whether they can talk through their own choices when you ask them to.

None of this requires new software — it's closer to what good writing teachers have always done, just applied more deliberately now that the stakes include an unreliable score in the mix. Staged drafts instead of one final submission. A quick conversation when something feels off, before any grade is affected. Judgment that stays with the teacher, with any automated signal treated as a prompt to look closer, never as the answer itself.

That's also the model Writing, no kidding is built around. Rather than generating a single cheating-probability score, it surfaces specific signals — tab switches away from the assignment, DevTools access, typing rhythm that doesn't match normal writing — for you to review yourself. Nothing gets auto-flagged as plagiarism and nothing gets auto-published as a verdict. See How to Run a Locked-Down Test Session for how that monitoring actually works day to day.

If you're building out a fuller academic integrity policy for AI-era writing assignments, it's worth pairing this process-based approach with clear expectations for staged drafts and check-ins — something we'll cover in more detail in an upcoming post. In the meantime, see How AI Grading Gives ESL Teachers Their Evenings Back for how the same "AI assists, teacher decides" principle plays out in grading.

---

See how Writing, no kidding flags signals without auto-verdicts. Start Free — no credit card required.