Guide
Do AI Detectors Flag Non-Native English Writers?
Do AI detectors flag non-native English writers? What research found, our own test on 879 essays by English learners, and how to read a result fairly.
Last updated Nov 15, 2026
In 30 seconds: yes, AI detectors can wrongly flag writing by non-native English speakers, and early research found very high rates. How often it happens depends on the detector and where it draws the line. In our own test of 879 essays by English learners, our free quick check called 1 essay AI-generated (0.1%) and left 2.8% as "can't tell". A wide inconclusive zone, not a single cut-off, is what keeps that number low.
| Our test, October 2026 | Essays | Called AI | Can't tell | Called human |
|---|---|---|---|---|
| Essays by English learners | 879 | 0.1% (1) | 2.8% | 97.0% |
| Human writing, mixed sources | 1,141 | 0.3% (3) | 0.6% | 99.1% |
Why non-native writing can look like AI
AI text detectors look for statistical patterns. Text written by a language model tends to use the most predictable word at each step, keep sentences even in length and follow familiar structures. Detectors learn to treat that predictability as a sign of AI.
Writing in a second language often shares the same surface features for good reasons:
- A smaller working vocabulary, so the same common words appear more often.
- Taught structures, such as five-paragraph essays, linking phrases and set expressions learned for exams.
- Simpler, more regular sentences, chosen to avoid mistakes.
None of this means the writer used AI. It means a detector that only measures predictability can confuse careful, formulaic human writing with machine output.
What research has found
The best-known study is from Stanford researchers in 2023 (Liang et al.). They ran seven popular GPT detectors on TOEFL essays written by non-native English speakers, and the detectors flagged more than half of those human essays as AI-generated. Essays by native-speaking US students were flagged far less often.
Detectors have changed since then, and results depend heavily on the tool and the threshold it uses. The finding still matters: a detector that has not been tested on second-language writing can be unfair to the people most likely to be suspected.
What our own test found
We tested the model behind our free quick check (Gradient, run on our own server) on two openly licensed collections of essays by English learners. We only count essays of 150 words or more, because the quick check does not give a verdict on shorter text.
| Collection | Writers | Essays (150+ words) |
|---|---|---|
| ArCC | Arabic-speaking high school, undergraduate and graduate students, 2016–2020 | 613 |
| SCoCESLE | Adult learners in Colombia (levels A2 to C1), mostly timed classroom essays, 2022 | 266 |
The result: 1 of the 879 essays was called AI-generated, a graduate student's essay with several spelling errors. 2.8% landed in the "can't tell" zone and 97.0% were called human-written.
The threshold is what makes the difference. We only call a text AI when the model's score is very high (0.90 or more). Scores between 0.54 and 0.90 are shown as "can't tell". If that middle zone were counted as AI, about 3–4% of these essays would have been flagged, and in our pre-release run about 18% of essays by the least advanced group (pre-sessional students in ArCC) scored in that zone. Lower proficiency means a higher chance of an inconclusive result.
The limits of our test
Our numbers are encouraging, but they are not a guarantee for every writer.
- Two collections only. Both are essays by non-native students. We found no commercially licensed collection of native-speaker student essays to compare against.
- Sample size. One false alarm in 879 essays gives a likely range of about 0.02% to 0.64%. Rare is not the same as never.
- Timing. A few SCoCESLE essays may have been written after ChatGPT's release in late 2022, although most were timed classroom work.
- One detector. These results apply to our free quick check. Other detectors, including our Deep check engine, were not part of this test.
The full method is on AI detector accuracy and limits.
How teachers should read a detector result
Treat any score as a reason to look closer, never as a verdict, and take extra care with students writing in a second language.
- Respect the middle zone. An "inconclusive" result is a real answer. It does not mean "probably AI".
- Ask for the process. Drafts, notes and version history in Google Docs or Word usually settle the question faster than any detector.
- Compare with earlier work from the same student, allowing for progress over a term.
- Talk to the student. People who wrote something can usually explain their choices.
Never use a detector result alone to accuse, grade or punish someone. Our guide on how to tell if text is AI generated covers the other checks in more detail.
What to do if your writing was flagged
If you are a student and a detector flagged your own work, you can usually show how you wrote it.
- Share your drafts and version history. They show the text growing over time.
- Ask which tool was used and what the score was. A result in a middle zone is not evidence of AI use.
- Offer to explain or rewrite a section in front of your teacher.
- Point to the research. The study above shows that detectors can misjudge second-language writing.
Frequently asked questions
Are AI detectors biased against non-native speakers? Some have been shown to flag non-native writing far more often than native writing. The risk depends on the detector and its threshold, so ask whether a tool was tested on second-language writing.
Does writing simply make my text look like AI? Plain, formulaic writing can score higher on some detectors. In our test, most simple essays still landed in the human or "can't tell" zones rather than being called AI.
Can I check my own essay before submitting it? Yes. Our AI text detector runs a free quick check on English text of 150 words or more. Read the result together with how our scale works.