AI Detectors and Non-Native English Writers: What the Research Shows

A practical guide for international students · Last updated 19 October 2026

If English isn't your first language, you may have a specific worry that most of your classmates don't: that your own, entirely human-written essay could get flagged as AI-generated simply because of how you write. That worry isn't paranoia — it's backed by real research.

What the research actually found

In 2023, a team led by Stanford professor James Zou tested seven widely used AI-detection tools against 91 essays genuinely written by non-native English speakers for the TOEFL exam. The results were stark: the detectors classified 61.22% of these authentically human essays as AI-generated. At least one detector flagged 97% of them. All seven detectors unanimously flagged 19% of them. By contrast, the same tools were near-perfect at correctly identifying essays written by native-English-speaking US eighth-graders as human-written.

The study's title says it plainly: "GPT Detectors Are Biased Against Non-native English Writers" (Liang et al., arXiv:2304.02819, 2023).

Why this happens

AI detectors don't read for meaning — they score statistical patterns, mainly perplexity (how predictable each word choice is) and burstiness (how much sentence length and rhythm vary across a passage). Writing that leans on simpler, more uniform sentence structures and a narrower vocabulary range — common in a second language, and often explicitly taught as "safe" academic style — scores lower on both measures. That's also, coincidentally, what AI-generated text tends to look like. The detector can't tell "this person is writing carefully in their second language" apart from "this was machine-generated." It just sees the same statistical signature either way.

What this means practically

A high AI-detection score is not proof of anything. Most detection providers, including Turnitin, now explicitly advise against treating a single score as conclusive on its own — this is exactly the kind of false positive they're warning about.
It's worth knowing this before you're accused, not after. If your work is ever flagged, you're in a stronger position if you can point to the actual research showing why that can happen to genuinely human writing from non-native speakers, rather than just asserting you didn't use AI.
It doesn't mean giving up on precise, formal academic English. The fix isn't writing worse — it's knowing the score itself isn't reliable evidence either way.

This is a known, published limitation of the technology, not a fringe complaint. If your university's academic integrity process treats an AI-detection score as decisive on its own, the Stanford study (and the providers' own guidance) is worth citing directly.

This is also exactly why LitMark never writes any part of your essay — it only finds real sources and checks your citations against them. Every word in your essay is still entirely your own, written in your own voice, which is the one thing no detector score can take away from you. See how it works.

The honest takeaway: AI-detection scores are a known, imperfect statistical guess, and the research shows they guess wrong more often against non-native English writers specifically. That's a limitation of the tool, not a reflection of your writing.

← Back to LitMark