Writing in English as a second language: why plain prose gets flagged and how to revise

If you write in English as a second or third language, you may have been told your work “looks like AI”, or you may fear it will be. The fear has a basis, and the fault is not in your English. Some detection tools score plain, careful prose as machine-like, and plain, careful prose is what many people write when they are working in a language they learned as adults. This article sets out what the research found, what that does and does not mean for you, and how to revise for the reader in front of you.

What the research found

The best-known test is from a Stanford team led by Weixin Liang. They gave seven detectors 91 essays that non-native English speakers had written for the TOEFL exam, and 88 essays by US eighth-graders. All the essays were human-written. The detectors were close to accurate on the eighth-graders. On the TOEFL essays, the average false-positive rate was 61.3 per cent, and 97.8 per cent were flagged as AI-written by at least one of the detectors. Liang et al., Patterns, 2023 (opens in a new tab).

The authors point to how these tools work. The methods they examine score how predictable a text is, a quantity called perplexity, and treat predictable text as machine-written. The same team checked abstracts from 1,574 accepted ICLR 2023 conference papers, submitted before ChatGPT existed, and found that abstracts by authors based in non-native-English-speaking countries had significantly lower perplexity than those by authors in native-English countries (in the preprint’s (opens in a new tab) analysis, p = 0.035). A writer with a smaller working vocabulary in a second language, using common words and steady sentence shapes, produces text these tools rate as predictable.

A separate study points to a related risk. Weber-Wulff and colleagues tested 14 tools and included documents that humans wrote in other languages and that were then machine translated into English. In their preprint (opens in a new tab), accuracy on those documents was about 20 per cent lower than on human-written English, and the average share that would have led to a false accusation rose from 2.4 per cent to 11.1 per cent.

Three cautions belong next to these figures. The Stanford samples were small, and the authors say so. Both studies date from 2023, and detection tools change without notice, so these numbers are not current error rates for any tool today. And vendors dispute the picture: Turnitin, for one, has stated its false positive rate is below 1 per cent (opens in a new tab). The mechanism is what has held up, and it is a reason to be careful, not a rate to quote.

What this does not mean

It does not mean your English is deficient, and it does not mean you should try to sound like someone else. The Conference on College Composition and Communication’s statement on second language writing and multilingual writers (opens in a new tab) asks instructors to assess organisation, development, word choice and grammar together, to look for what is rhetorically effective in a text and not only for surface errors, and to treat multilingual writers’ linguistic resources as strengths. Plain prose is also what the federal plain-language guide (opens in a new tab) asks for.

It also does not mean that changing your style will protect you. We cannot say what any detector will report about any text, before or after a revision, and we would distrust anyone who says they can. So revise for your reader, and protect yourself with evidence of how you worked: dated drafts, notes and version history. That is what the evidence checklist covers, and it is worth building before anyone asks.

Revising for readers

These are ordinary writing-centre moves. They make a paper more interesting to read, and they suit any writer, in any language.

1. Add what only you know. The most general sentence in a paper is often the weakest. An example from your country’s school system, your own lab or your job says something a generic sentence cannot, and it is yours. One specific detail per section is a good target.
2. Vary the rhythm. Steady sentence lengths tire readers in every language. Put a short sentence after a long one when the point matters. The writing centre handouts we have linked in fixing your own stiff draft show how.
3. Use verbs. If your sentences are built from long noun phrases, turn some back into actions: “the analysis of the data was carried out” becomes “we analysed the data”. This is good advice in any language.
4. Borrow structure, not sentences. The University of Manchester’s Academic Phrasebank (opens in a new tab) collects example phrases organised by the parts of an academic paper and by what the phrase does. Use it to see how a writer signals a contrast or states a limitation, then say your own point in your own words. Strings of stock phrases are one of the things that make prose feel flat.
5. Read aloud, then ask a person. Where you stumble, a reader would too. If your university has a writing centre, ask whether it has tutors who work with multilingual writers; a tutor can show you a pattern once and you can then apply it yourself.

If you draft in your first language

Many people think through a paper in the language they are most comfortable in, then translate. That is a legitimate way to work, and your course may treat it as a use of a tool that needs a note. The Modern Language Association’s guidance on generative AI lists translating words among the functional uses (opens in a new tab) a writer should acknowledge. Ask your instructor which tools count in that course, get the answer in writing, and keep your first-language draft: it is strong evidence of authorship. For a message you can send, see how to read your course’s AI policy.

If your course allows a rewriting tool and you choose to use one, the rewriter on the home page shows the original beside the rewrite with every changed word marked. Treat it like any editor’s suggestions: it may change wording, order and structure, and can also change a number or drop a qualifier, so check the result against what you meant. Here is how to check it.

If you are the instructor

A flag on a second-language writer’s paper is a reason for a conversation and for corroboration, not a finding. There is more on how to run that conversation fairly in talking to a student about a suspected AI paper.