Are AI detectors biased against non-native English speakers?
Short answer
Older detectors were, badly. In a 2023 Stanford study, seven popular detectors labeled 61.3% of essays written by non-native English speakers for the TOEFL exam as AI-generated on average, while they classified US eighth-graders' essays accurately.
Newer detectors do much better on plain human writing: in a 2026 study of theses by non-native graduate students, Turnitin, Copyleaks and Pangram reported 0% AI. But the bias hasn't disappeared. When non-native writers polish their text with AI, at least one widely used detector gives it higher AI scores than native writers' polished text, so any flag on a non-native writer's work deserves extra care.
What did the Stanford study find?
Researchers at Stanford (Liang, Yuksekgonul, Mao, Wu and Zou) ran human-written essays through seven widely used detectors: Originality.ai, Quil.org, Sapling, OpenAI's GPT-2 output detector, Crossplag, GPTZero and ZeroGPT. Turnitin was not tested. The study was published in Patterns in July 2023.
| Essays tested | Written by | Labeled AI-generated (average of 7 detectors) |
|---|---|---|
| 91 TOEFL essays | Non-native English speakers | 61.3% |
| 88 US eighth-grade essays | Native speakers | Classified accurately (5.19% in the preprint) |
| The same 88 essays, word choice simplified by ChatGPT "as if written by a non-native speaker" | Native speakers, edited | 56.65% (preprint) |
| The 91 TOEFL essays, word choice enhanced by ChatGPT to sound native | Non-native speakers, edited | 11.6% |
- All seven detectors agreed on 19.8% of the human TOEFL essays, and at least one detector flagged 97.8% of them.
- The cause, according to the authors: detectors rely on how predictable the wording is (perplexity). Writers with a smaller vocabulary and less varied sentences look more predictable, so they look more like AI.
Is it still true in 2026?
On plain, unedited writing, newer detectors are far better. On edited writing, the picture is mixed:
- Unedited human text, 2026: In a study published in June 2026 by researchers at Vrije Universiteit Brussel, 40 theses written in English by non-native graduate students before ChatGPT existed got 0% AI from Turnitin, Copyleaks and Pangram. GPTZero showed a small false positive rate.
- Same essays, newer tool: A 2026 replication ran the commercial detector Plagramme on the Stanford TOEFL essays. It wrongly flagged 23.1% of them and 0% of the native essays. That's progress since 2023, but the gap is still there.
- AI-polished text, 2025: A PeerJ study found GPTZero accurate on purely human abstracts (0% false positives). But on abstracts polished by ChatGPT or Gemini, it gave non-native authors significantly higher AI scores: a mean of 44.61% versus 30.68% for native authors. It labeled 25% of their polished abstracts 100% AI, versus 11% for native authors.
- AI-polished research letters, 2025: In JAAD International, GPTZero classed 97 to 100% of original letters by authors from China, Korea, Japan and the US as human. After ChatGPT polished them, 75 to 85% were flagged as AI.
- Professional human editing, 2026: A study of 135,389 non-native manuscripts and their professionally edited versions found that the same edits raised AI scores in some research detectors and lowered them in others. It tested 13 research detectors, not commercial tools.
What do the detector companies say?
These are the companies' own tests, not independent results:
- Turnitin (October 2023): no statistically significant difference in false positive rates between English learners and native writers for documents of 300+ words, with a larger gap for 150 to 300-word texts. Turnitin points out that the Stanford TOEFL essays were all under 150 words.
- GPTZero (October 2026): its 4o model labeled all 914 essays by English language learners in a public test set as human.
- Pangram (July 2026): Pangram 4 had one false positive in 24,586 texts by English learners (0.0041%).
Why does it matter where the bias comes from?
Because it changes what non-native writers should watch for. With 2023-era detectors, plain human writing was the problem. With current detectors, the bigger risk is human writing that went through an AI rewriting or polishing step. Then the detector sees real AI involvement, and in the 2025 PeerJ study GPTZero scored that involvement higher for non-native writers.
Advice for non-native English writers
- Keep your own draft and version history before any editing step. It shows what you wrote yourself.
- Spelling and grammar fixes are usually fine. Be careful with features that rewrite whole sentences or "improve fluency."
- If your course allows AI help with language, follow its disclosure rules and keep the chat.
- Check your final text before you submit it, and look at which passages get flagged. How to check your essay.
- If you're flagged, say clearly that English is not your first language and bring your drafts. What to do after a flag.
Advice for teachers
- Bias against non-native writers is one of the reasons Vanderbilt (2023) and Waterloo (2025) stopped using Turnitin's AI detection.
- Compare a flagged text with the student's earlier work, not with native-speaker norms.
- Ask about the student's process and any language tools they used before drawing conclusions. A fair process for checking student work.
Questions people also ask
Do AI detectors flag ESL students more often?
The 2023 detectors did: seven of them flagged 61.3% of TOEFL essays on average. Current tools flag far fewer unedited human texts, but a 2025 study found GPTZero gave AI-polished abstracts by non-native authors higher AI scores (mean 44.61%) than polished abstracts by native authors (30.68%). Edited text is where non-native writers are most exposed today.
Does using a translator or grammar checker make my writing look AI-generated?
Spelling and grammar fixes usually don't; Pangram, for example, defines light copyediting and literal translation as human writing. Rewriting whole sentences with an AI tool is different: studies in 2025 and 2026 found that AI-polished human text was often flagged. Keep your own draft so you can show what you wrote.
Which AI detector is least biased against non-native writers?
No tool is proven bias-free. In an independent 2026 study, Turnitin, Copyleaks and Pangram reported 0% AI on 40 theses by non-native graduate students, while GPTZero showed a small false positive rate. GPTZero, Pangram and Turnitin all publish their own tests claiming no meaningful bias, but those are vendor results.
Sources
- Liang et al., GPT detectors are biased against non-native English writers, Patterns 4(7) (Jul 2023)
- Liang et al., preprint with methods and detector list (arXiv 2304.02819)
- Van Vlasselaer et al., International Journal for Educational Integrity 22, article 16 (Jun 2026)
- Pratama, GPTZero and AI-assisted abstracts by native and non-native authors, PeerJ Computer Science (Jun 2025)
- Wang et al., AI polishing and AI detection of research letters, JAAD International (Jun 2025)
- Al Ali, Helcl and Libovický, bias of AI detectors on non-native writing (arXiv, Feb 2026)
- Park, Jeong and Kim, Style as a Confound (arXiv, Aug 2026)
- Turnitin: AI detector shows no statistically significant bias against English language learners (Oct 2023)
- GPTZero: no bias against ESL writers (Oct 2026)
- Pangram 4 technical report (arXiv, Jul 2026)
- Vanderbilt University: why we're disabling Turnitin's AI detector (Aug 2023)
- University of Waterloo: Discontinuing use of AI detection functionality in Turnitin