Flagged by an AI Detector as an ESL Student? A Practical Guide

AI writing detectors do not fail evenly. Research suggests they can be more likely to misread writing by people working in a second language. If you are an ESL student and an AI flag landed on work you wrote yourself, this guide explains why that can happen and what to do next.

The main evidence is a 2023 study by Stanford researchers (Liang et al., published in Patterns). Seven detectors available at the time labeled, on average, 61% of 91 human-written TOEFL essays as AI-generated, while essays by US eighth-graders were classified far more accurately (Liang et al., 2023). The study tested specific tools on a small set of essays, so it does not tell you the error rate of the detector your school uses. It does show that this kind of mistake is real and documented. We cover the study in detail in what the Stanford study actually found.

Why detectors can get this wrong

Detectors do not know how a text was written. They look for statistical patterns in the finished text. One common signal is perplexity: how predictable each word is to a language model. Language models tend to generate predictable text, so many detectors treat predictable text as a sign of AI.

When you write in a language that is not your first, you often choose words you are sure of. That is a smart way to avoid mistakes. It also makes your text more predictable. In the Stanford study, when the researchers enriched the vocabulary of the same TOEFL essays, the detectors' average false-positive rate dropped sharply. Same writers, same ideas, different words, different results.

A second common signal is burstiness, or how much sentence length and structure vary. Writers who build each sentence carefully, using patterns they have practiced, can produce a steady rhythm. To a detector, steady rhythm can look machine-like. To a human reader, it often just looks careful.

It is not only ESL writers

The same mechanism can affect anyone whose writing is predictable for innocent reasons: technical writers using standard terminology, students writing in a formal register they were taught, or anyone whose draft has been heavily polished. A detector score describes how the text looks to that tool. It is not a finding about who wrote it.

What to do if you are flagged

• Bring your drafts and version history. Google Docs (File → Version history → See version history) and Word files saved to OneDrive or SharePoint keep revision history. A record of the essay developing over days, with edits and false starts, is strong supporting evidence of your process.
• Ask what the report shows. Some tools highlight the specific passages they scored as AI-like. Turnitin, for example, shows no percentage or highlights for scores between 1% and 19%, because its testing found more false positives in that range (Turnitin guides). Knowing which passages were flagged lets you explain how you wrote them.
• Talk to your instructor. Many concerns can be discussed directly. Explain how you planned, drafted, and revised, and offer your notes. If your institution has a formal process, read the policy, ask what support is available (an advisor, the international student office, or the dean of students office), and respond in writing.
• Point to the vendor guidance if needed. Turnitin's own guidance says its AI score may misidentify text and should not be the sole basis for adverse action against a student.

Should you edit your own writing just to lower a score?

Some students ask whether they should change their writing so detectors stop flagging it. You wrote the essay. You should not have to disguise your own work, and no editing step, by hand or with a tool, can guarantee a detector result.

What does help is ordinary good revision: read the essay aloud, fix sentences that feel stiff, and use the most precise word you know when you are confident it is right. That improves the essay for your reader, which is the point.

If your course allows AI-assisted editing and you want suggestions on phrasing, HumaText can propose a phrasing pass on a short section. Read every change and keep only what still sounds like you and says what you meant. Its writing-pattern score is a rough internal heuristic, not a prediction of what Turnitin or any other detector will report. Whatever you do, keep your version history first. It matters more than any score.

Where this goes

Detector vendors keep updating their models, and some report that their tools do not show this bias. Turnitin, for example, published 2023 testing that found no statistically significant difference in false-positive rates for English-language-learner writing on documents above its 300-word minimum (Turnitin). That is a vendor testing its own product. Weigh vendor results alongside independent research, and treat any single score as one signal, not proof.

If you were flagged and you did not use AI, you are not imagining the problem. Bring your evidence, ask what was flagged, and talk to your instructor.