If English isn't your first language and you've ever been nervous about AI detectors flagging work you genuinely wrote, that fear isn't paranoia — it's the documented finding of peer-reviewed research. The Stanford study "GPT detectors are biased against non-native English writers" tested widely used detectors on essays by non-native writers and found they misclassified more than half of them, on average, as AI-generated — while performing far better on essays by native speakers. This post is about why that happens mechanically, and what you can do about it.
Why it happens: the perplexity trap
AI detectors measure how predictable text is — word-by-word probability, sentence-rhythm evenness. Language models produce predictable text; that's the signal. But writing in a second language pushes humans toward the same statistical territory for entirely human reasons: a working vocabulary that favors common words, grammar constructions you're confident in and reuse, sentence patterns learned from textbooks, and the natural caution of not wanting to make errors. Safe, correct, conventional prose — which is exactly what high-probability machine text looks like to a classifier. The detector isn't measuring honesty; it's measuring linguistic risk-taking, and second-language writers rationally take fewer risks.
What Turnitin (and universities) say about it
Turnitin has said it trained on non-native English writing and disputes that its detector shows this bias at meaningful levels; the company also masks scores under ~20% partly for false-positive reasons. Independent evidence, meanwhile, was strong enough that some universities disabled AI detection entirely, citing exactly this population's exposure (the accuracy evidence, laid out). The honest summary: the risk is real, its exact size on any given detector version is unknowable from outside, and international students carry more of it than anyone else.
Protecting yourself without changing who you are
- Keep receipts as a default habit. Draft in Google Docs or Word with version history on. A timestamped trail of the essay growing across sessions is the evidence that ends most false-flag conversations quickly.
- Let your specifics in. The strongest anti-flag signal is content no model would produce: your course's readings, the seminar discussion, your own examples. Specificity is unpredictability, and it's also just better writing.
- Be careful with heavy polishing tools. Aggressive grammar-tool rewriting smooths out the remaining human texture — accept fixes, not voices.
- Know your number before submission. This is the group for whom pre-checking makes the most sense: a real Turnitin check shows you the AI report in advance, so a false flag becomes a revision task instead of a hearing. If it flags passages you wrote, rework their rhythm — vary sentence lengths, break the formula — and re-check.
And if you're reading this as an instructor
The research is one more reason Turnitin's own guidance says the AI indicator shouldn't be the sole basis for action. If your flagged student is writing in their second language, the flag is statistically weaker evidence than it looks — ask for drafts and version history before drawing conclusions. The students most likely to be wrongly flagged are often the least equipped to argue back.