AI detectors systematically flag non-native English writers
The tools institutions use to catch AI-written work misclassify second-language writing as machine-generated. If you write clear, plain English, this is your problem too.
If you are a student, you have probably been told your work may be run through an AI detector. It is worth knowing what those tools actually do, because the answer affects people who have done nothing wrong.
Liang et al. (2023) evaluated several widely-used GPT detectors and found they systematically misclassify writing by non-native English speakers as AI-generated. In their evaluation, more than half of non-native student essays were flagged, while native-speaker writing was classified correctly far more often.
That is not a rounding error. That is a tool that produces different outcomes for different groups of students, deployed in decisions with real consequences.
Why it happens
Detectors largely key on text perplexity — roughly, how predictable the word choices are. Machine-generated text tends to be statistically smooth: common words, conventional phrasing, low variety.
So does competent second-language writing. Someone writing carefully in a language they learned later tends to use a more constrained vocabulary and more conventional constructions. Not because the writing is worse — because that is what careful non-native writing looks like.
The detector cannot tell the difference, because at the level it measures, there is no difference.
It also means the same risk applies to plenty of native speakers: if you write plainly, structure things clearly, and avoid flourishes, you look more like the thing being detected.
The other half
Weber-Wulff et al. (2023) tested detection tools from the other direction and found accuracy well below what consequential decisions would require, degrading further when text was lightly edited or paraphrased.
So the picture is:
- False negatives are easy to produce. Anyone trying to evade detection can.
- False positives fall hardest on non-native writers.
A tool that misses the people gaming it and flags the people who are not is worse than no tool, because it carries the authority of a number.
What to actually do
Not "avoid AI" — that does not help, because you can be flagged for writing you did entirely yourself.
Keep your drafts. Write in something with version history — Google Docs, or git. A document that grew over two weeks looks completely different from one that appeared complete in a single paste, and that history is nearly impossible to fabricate afterwards.
Keep your notes. Handwritten plans, reading notes, the outline you abandoned.
Keep a short use log. One line per use: what you asked, what you did with it. Thirty seconds each, and if you are ever questioned it is a far stronger answer than insisting.
Read your institution's actual policy, not a rumour about it.
None of this is paranoia. It is the same evidence you would want anyway, and the log has a second benefit: it is hard to write "asked it to write my conclusion" in a document you might one day show someone.
The full treatment, including where the line actually sits, is in academic integrity, honestly.
All notes