Why detectors read second-language English as machine writing
Most detectors score text on how predictable it is. The technical term is perplexity: if each word is close to the word a language model would have guessed next, the passage scores as machine-like. Variation in sentence length, sometimes called burstiness, gets weighed the same way.
That measurement punishes one kind of writer in particular. Someone working in a second language tends to use a narrower and safer vocabulary, reach for constructions they are confident are correct, and keep sentence structure regular. Those are the habits of a person being careful. To a perplexity model they are indistinguishable from generated text.
The size of the gap is the part worth sitting with. A 2023 study from Stanford found AI detection tools flag 61% of writing by non-native English speakers as AI-generated. For native speakers the false positive rate drops to around 3%. Same tools, same task, and the error rate multiplies roughly twentyfold depending on who wrote the essay.