If English is your second language, a detector is far more likely to flag your own work. The cause is how the tools are built, and it says nothing about how you write.
Last reviewed 12 November 2026
The best-known study on this comes from a Stanford group, published in the journal Patterns in 2023. The researchers took 91 essays written for the TOEFL English exam by non-native speakers and ran them through seven widely used AI detectors. They did the same with essays by US eighth-grade students.
All of those essays were written by people. The detectors were roughly twelve times more likely to call a human essay machine-written when its author was writing in a second language.
Most detectors lean on how predictable the wording is. Text built from common words and standard sentence patterns looks, statistically, like the output of a language model, which also favours common words and standard patterns.
Careful second-language writing has the same surface features for an entirely different reason. Writers working in a second language tend to:
Those are sensible choices. They are also exactly what the statistics reward with a high "AI" score. The Stanford group confirmed the mechanism: when the same essays were reworded with more varied vocabulary, far fewer of them were flagged. The ideas and the authors had not changed, only the range of words.
A high score on second-language writing mostly measures vocabulary range. It is weak evidence about who wrote the text, and for this group it is close to no evidence at all.
More on course-level policy is in AI detection for educators.
The study tested tools available in 2023, and vendors have changed their products since. The underlying issue has not gone away: any method that treats predictable wording as a sign of machine writing will be harder on writers with a narrower active vocabulary. Before trusting a newer tool, ask for independent results on second-language writing specifically. A general accuracy figure does not answer that question, for the reasons set out in What a 1% false-positive rate actually means.
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7). arXiv:2304.02819