Home / Guides

Are AI detectors biased against non-native English speakers?

Short answer: yes, and the numbers are hard to argue with. In a Stanford study, seven popular AI detectors wrongly labeled about 6 in 10 essays written by non-native English speakers as AI-generated, while they barely misfired at all on essays by native English speakers. If English is not your first language, your honest writing is far more likely to be treated as a robot's work.

Updated November 2, 2026 · By the IsThisAI? team

The study that proved it

In 2023, researchers at Stanford (Liang, Yuksekgonul, Mao, Wu and Zou) tested seven widely used GPT detectors on two sets of human-written essays: 91 TOEFL essays by non-native English speakers, collected from a Chinese forum, and 88 essays by US eighth graders from the Hewlett Foundation's ASAP dataset. Not one essay in either set came from a machine.

On average, the detectors called 61.3% of the TOEFL essays AI-generated. The eighth-grader essays were classified almost perfectly. At least one of the seven detectors flagged 97.8% of the TOEFL essays, and all seven agreed that 19.8% of them were AI-written. None of them were.

The researchers then ran a revealing experiment. They asked ChatGPT to "enhance the word choices" of the TOEFL essays "to sound more like that of a native speaker." The false positive rate dropped from 61.3% to 11.6%. Then they did the reverse: they had ChatGPT simplify native essays "as if written by a non-native speaker." Misclassification shot up. The bias is not about who the writer is. It is about how the writing sounds (Liang et al., 2023).

Why it happens: predictable writing looks "AI-like"

Most detectors measure something called perplexity: how surprised a language model is by each word in your text. Turnitin describes its own logic in plain terms: language models tend to pick the next word in a "consistent and highly probable fashion," while human writing "tends to be inconsistent and idiosyncratic" (Turnitin FAQ).

That logic has a hole. Writing in a second language is often careful, simple and consistent. You use the words you are sure about and build sentences you know are correct. That is good writing discipline, but to a detector it reads exactly like a machine picking the most probable next word. The paper puts it bluntly: "The design of many GPT detectors inherently discriminates against non-native authors, particularly those exhibiting restricted linguistic diversity and word choice." Non-native speakers are known to show less lexical richness, syntactic diversity and grammatical complexity, which are real traits of learner writing and also the exact traits detectors punish.

It is not just school essays

The same study looked beyond students. The researchers analyzed papers from ICLR 2023, a major machine learning conference, and found that papers by first authors from countries where English is not the native language showed lower text perplexity than papers by native-speaking authors, even after accounting for review ratings. So the fingerprint shows up in academic publishing too. Anywhere a detector is used to police AI use, from conferences to hiring screens, non-native writers carry extra risk.

Then there is the cruel twist. Because detectors punish simple writing, the authors argue that non-native writers may feel pushed to use AI precisely to sound more "native" and evade detection. They also found that asking ChatGPT to "elevate the provided text by employing literary language" dropped detection rates on AI-generated essays to near zero. So these tools fail in both directions: they wrongly accuse human non-native writers, and they can be fooled by one well-crafted prompt. The students most at risk of a false accusation are also the ones with the strongest incentive to actually use AI, which is the opposite of what detectors are supposed to achieve.

What to do if your writing gets flagged

First, do not panic. A flag is a probability, not proof. Here is what actually helps:

1. Keep your drafts. Version history in Google Docs or Word, outlines, notes, even messy first drafts. They are the best evidence that the words are yours.
2. Show your process. If you are accused, walk the instructor through how you wrote the piece: the sources, the revisions, the parts you struggled with.
3. Check yourself first. Paste your text into a free detector before you submit, so a surprise flag does not catch you off guard.
4. Speak up. Tell your instructor that detector bias against non-native writers is documented research and point to the Stanford study. Most teachers have never heard of it.
5. Do not reach for a humanizer. Paraphrasing tools sold as ways to beat detectors are unreliable and can turn a false accusation into a real integrity problem.

For the full evidence-gathering playbook, see how to prove you wrote it yourself.

What teachers and schools should do

The Stanford authors are explicit: they caution against using GPT detectors in evaluative or educational settings, "particularly when they may inadvertently penalize or exclude non-native English speakers." Turnitin itself says its AI score should not be used as the sole basis for action against a student.

If you teach or set policy, treat detector output as a starting question, never an answer. Ask for drafts. Use assignments that show the writing process, not just the finished product. And remember where the errors land: not randomly, but on international students, immigrants and anyone writing in a second language. Using these tools uncritically turns a statistical bias in a piece of software into a punishment that falls on one group of students. That is not academic integrity. It is the opposite of it.

More guides

Sources

1. Liang, Yuksekgonul, Mao, Wu & Zou, "GPT detectors are biased against non-native English writers," Patterns 4(7), 100779 (2023). https://pmc.ncbi.nlm.nih.gov/articles/PMC10382961/, preprint: https://arxiv.org/abs/2304.02819v1
2. Turnitin, "Turnitin's AI writing detection capabilities FAQs" (Turnitin Guides, updated Oct 7, 2026). https://www.turnitin.com/products/features/ai-writing-detection/faq
3. Weber-Wulff et al., "Testing of detection tools for AI-generated text," International Journal for Educational Integrity 19, 26 (2023). https://link.springer.com/article/10.1007/s40979-023-00146-z
4. IsThisAI? internal benchmark of six open-source detectors on 142 documents, including 13 AI-written resumes (Oct 2026). https://getisthisai.com/about/

This guide is general information, not legal or academic-policy advice. AI detection results, including ours, can be wrong.