Why detectors flag non-native English writers

What the research actually found

In 2023 a group of Stanford researchers, Weixin Liang and colleagues, ran 91 TOEFL essays written by non-native English speakers through seven widely used AI detectors. The paper appeared in the journal Patterns under the title 'GPT detectors are biased against non-native English writers'.

The detectors misclassified a large share of those human-written essays as machine-generated, on average more than half. Nineteen of the 91 essays were flagged by all seven detectors at once. The same detectors, given essays written by US eighth-graders, correctly called them human almost every time.

The researchers also showed the mechanism directly. When they asked a language model to rewrite the TOEFL essays using richer, more varied vocabulary, the misclassification rate collapsed. The detectors were not detecting machines. They were detecting limited vocabulary.

Why second-language writing produces the same signal as a model

Detectors mostly score two things: how predictable each word is given the ones before it, and how much sentence length and complexity vary across a passage. Machine prose is predictable and even. So is careful second-language prose, for reasons that have nothing to do with machines.

• You choose words you are sure of. Safe, common vocabulary is exactly what a language model also picks, because both of you are optimising for being correct.
• You avoid idiom, slang and wordplay: the surprising, low-probability choices that read as human to a detector.
• Exam preparation teaches templates. IELTS and TOEFL coaching drills fixed openings, connectives and paragraph shapes, which flattens variation by design.
• Sentence rhythm carried over from your first language is often more regular than native English rhythm.
• You proofread harder than a native speaker does, and every revision pass smooths out more of the irregularity that reads as human.

The translation problem, stated honestly

If you drafted in your own language and ran it through a translation tool, the words that were submitted really were produced by a machine, even though the thinking was entirely yours. Detectors score translated text high, and you should expect that.

This is a different conversation from the one above and it is worth separating clearly. Say what you did: you wrote the argument yourself in your first language, and used translation software to render it in English. Bring the original. A full draft in your own language is strong evidence of authorship, and it is the artefact most students in this situation forget to offer.

Whether translation is permitted is a policy question at your institution, not a detection question. Ask, and ask in writing.

How to raise this without it sounding like an excuse

The failure mode here is leading with unfairness. It puts the instructor in the position of defending both the tool and the institution, and it moves the conversation away from the only thing that actually helps you, your process.

Put it second, in one sentence, framed as context rather than defence. Then go straight back to your drafts.

What to ask your institution

• Does the academic integrity policy say a detector score alone is insufficient evidence? At many institutions it does, in writing.
• Is second-language status recognised anywhere in the policy as a factor to weigh?
• Can the writing centre, language support office or your ESL instructor confirm your level and your appointment history? Records of tutoring sessions on this specific assignment are excellent evidence.
• Is there an international student adviser whose job includes exactly this? There usually is, and they have handled the same case several times this term.

What makes it worse

• Running the text through a rewriting tool to 'fix your English' before submitting. Heavy machine rewriting is the one thing that genuinely does produce machine-patterned text, and it is hard to explain afterwards.
• Accepting every rewrite suggestion from a grammar tool rather than its corrections. Corrections fix errors; rewrites replace your sentences with the tool's.
• Deleting the first-language draft or the translation history to look tidier. That is the evidence.
• Leading with discrimination language before you have shown any of your process. It may be substantively fair and it still ends the collaborative version of the conversation.

If you are still studying, build the record now

Second-language writers carry more risk per assignment than anyone else, so the cheap protections are worth more to you than to your classmates.

• Draft in Google Docs or Word online, where revision history is recorded automatically.
• Keep your notes and outlines in your first language and do not delete them.
• Book writing centre appointments and keep the confirmation emails.
• If you translate, keep the source document and be ready to say so before anyone asks.