AI Detection and ESL Students: Are Non-Native Writers Unfairly Flagged?

Sometimes, yes: in the best-known study, seven AI detectors labeled 61% of essays by non-native English speakers as AI-written while scoring US students' essays almost perfectly. The evidence is mixed, though: Turnitin, which that study did not test, reports no significant bias for documents of 300 words or more, and newer research finds the effect depends on the detector, the language and how polished the writing is. Either way, a detector score alone is weak evidence, and the best protection for an ESL writer is being able to show how the work was written.

Below is what each side of the debate actually measured, with links to the sources, followed by practical steps.

The study behind the headlines: Liang et al. (2023)

Stanford researchers led by Weixin Liang and James Zou posted GPT detectors are biased against non-native English writers in April 2023. It was published in the journal Patterns in July 2023.

What they tested. Seven public detectors, accessed on 15 March 2023: Originality.ai, Quil.org, Sapling, OpenAI's detector, Crossplag, GPTZero and ZeroGPT. Turnitin was not included. The human writing was 91 TOEFL essays from a Chinese educational forum and 88 essays by US eighth-graders.

What they found.

• The detectors were near-perfect on the US essays.
• On the TOEFL essays, the average false positive rate was 61.22%.
• All seven detectors flagged 18 of the 91 TOEFL essays, and at least one flagged 89 of them (97.80%).
• When ChatGPT was used to make the TOEFL essays' word choices sound more native, the average false positive rate fell to 11.77%.
• When ChatGPT simplified the US essays' word choices "as if written by a non-native speaker", the misclassification rate rose from 5.19% to 56.65%.

Those figures come from the preprint; the journal version reports some of them slightly differently (61.3% and 11.6%). The authors' advice in the journal version was direct: "we strongly caution against the use of GPT detectors in evaluative or educational settings, particularly when assessing the work of non-native English speakers."

Why non-native writing can look "predictable"

Most detectors estimate how predictable a text's word choices are, a measure called perplexity (our guide to how AI detection works explains it). Language models favor common, probable words. So does a careful writer working in a second language: familiar vocabulary, safe sentence patterns, transitions learned from textbooks.

Liang's team found evidence for this link. The TOEFL essays that every detector flagged had significantly lower perplexity than the rest, which suggests detectors "may penalize non-native writers with limited linguistic expressions." In a separate check, abstracts of ICLR 2023 papers by authors based in non-native English-speaking countries also had lower perplexity than those by authors in native English-speaking countries.

Turnitin describes its own model in similar terms, as trained on differences in "word probability" (FAQ). And The Markup reported in August 2023 that Turnitin's chief product officer "admitted the tool ends up learning that more complex writing is more likely to be human."

The counter-evidence

Turnitin: no significant bias at 300 words or more

On 26 October 2023, Turnitin's principal machine learning scientist David Adamson published the company's own test. Turnitin sampled up to 2,000 texts for each combination of English learners ("L2") or native writers ("L1") and length, either 150 to 300 words or 300 words and over, drawn from three open datasets (ASAP, PELIC and ICNALE).

• For documents of 300 words or more, the difference in false positive rates between English learners and native writers was "small, and not statistically significant." Each group's rate was "slightly higher than our overall target of 0.01 (1%)" but not significantly different from it (Turnitin).
• For documents of 150 to 300 words, the gap between the groups was larger and "significantly greater than our target" (same post).

The conditions matter. Turnitin says most of the English-learner documents were "short persuasive or informative writing tasks", and that "a comparable collection of publicly available full-length university level documents from L2 and L1 writers was not available." It also criticized Liang's study as resting on 91 TOEFL practice essays, "all of which are less than 150 words long", which is below the length Turnitin will score. Turnitin requires at least 300 words of prose before it produces an AI report at all (FAQ). This is a vendor testing its own product, but it did publish its method.

Originality.ai's rebuttal (a vendor claim)

Originality.ai's founder and CEO Jonathan Gillham wrote a rebuttal of the Stanford study (the page is dated 11 December 2025). It argues the study was flawed because of its 91-essay sample, the comparison with US eighth-graders, and an out-of-date detector: the study used Originality.ai's version 1.1. The company then ran its own detector on "over 1,500 essay samples" of IELTS essays and reports that it "accurately identified 1,526 as human-written and incorrectly labeled only 81 as AI-generated", a false positive rate of 5.04%.

Read that as what it is: a vendor's test of its own product, described in a blog post. The same post says "Originality.AI is not for academic use" and that its tool is built for publishers and marketers.

ETS (2024): carefully built detectors need not be biased

Researchers at ETS published Detecting ChatGPT-Generated Essays in a Large-Scale Writing Assessment: Is there a Bias Against Non-Native English Speakers? in Computers & Education in August 2024. The full article is not open access. Co-author Jiangang Hao later wrote that this study and a companion one, "based on a large dataset, showed that carefully developed detectors do not necessarily bias against non-native English speakers" (Hao, ETS Research Institute). Note the wording: these were detectors the researchers built for an assessment setting, not the off-the-shelf tools in Liang's study. Hao also warns that "some detectors may exhibit bias against certain demographic groups" and need careful evaluation before any consequential decision.

Czech (2026): no systematic bias in a different language

A February 2026 study from Charles University and the University of Oslo, Different Time, Different Language (accepted to the EACL 2026 Student Research Workshop), revisited the question in Czech with 2025-era models and detectors. Essays by non-native speakers of Czech were not lower in predictability measures than native essays, and detectors from three families, including one commercial detector, showed no systematic bias against non-native writers. The authors conclude that the bias "is language dependent." Their proficient non-native sample was small, 29 essays alongside 450 from moderately proficient writers.

The newest twist: editing moves scores both ways

The largest study so far is Style as a Confound, submitted on 27 August 2026 and presented at EMNLP 2026. It compared 135,389 manuscripts by non-native English writers with the same manuscripts after professional native-speaker editing. On human-written documents from 2018 to 2022, before ChatGPT, the 13 detectors' false positive rates ranged from 0.0% to 100.0%. The same edits pushed some detectors toward "AI" and others toward "human", and the effect grew with the amount of editing.

For ESL writers, the lesson is uncomfortable: both plain, careful English and heavily polished English can move a score, and which way depends on the detector. The authors conclude that detector outputs "should not be treated as definitive evidence of AI use, especially in academic settings where editing is common." The study tested publicly available research detectors, not Turnitin or GPTZero, and one co-author works at the editing company whose records were used.

So, is there bias? The evidence side by side

The fairest reading: the problem Liang found was real for the tools and short essays tested in 2023; some vendors report that their own products do not show it (Turnitin, for documents of 300 words or more); and newer research shows detector behavior depends heavily on the tool, the language and the style of the text. Several of these sources, including Turnitin itself, say a score should not be the sole basis for action against a student.

What it looks like in practice

The Markup's reporting gives one detailed example. At Johns Hopkins, lecturer Taylor Hahn saw Turnitin label "more than 90 percent" of a student's paper as AI-generated; the student immediately showed him drafts and "PDFs with highlighter over them", and he concluded the tool was wrong. Over the semester he noticed Turnitin was much more likely to flag international students' writing (The Markup). That is one instructor's observation, not a study, but it matches why the research matters. More cases are collected in our explainer on AI detection false positives.

Practical steps for ESL writers

1. Keep drafts and version history

Write in a tool that records versions. In Google Docs, version history sits behind the "Last edit" button, and naming a version stops Google from merging it (Google help). In Word, version history only works for files stored in OneDrive or SharePoint (Microsoft support). Keep your outlines, notes and sources too. Our guide on how to prove your writing is human has the full evidence checklist.

2. Disclose permitted tool use

If your course allows a translator, a grammar checker or an AI tool, note what you used and where, and keep your text from before the tool changed it. A University of New Hampshire handout on documenting your writing process puts it simply: "Transparency is always safer than silence."

3. Ask which detector is used, and check its limits

Detectors differ, so ask which one your school uses. For Turnitin: it needs at least 300 words of prose, it shows scores from 1% to 19% only as *%, and it scores only long-form English, Spanish, Japanese and Modern Arabic (FAQ). Students cannot see the AI report, but instructors can download it and share it (same FAQ; more in can students see their Turnitin AI score?).

4. Be careful with tools that rewrite heavily

Liang's experiments show that changing word choice alone can swing detector results in either direction. The UNH handout lists grammar tools "set to rewrite extensively" and "humanizing" tools among those to avoid unless your instructor allows them, because they can erase your visible drafting history. That includes our own AI humanizer for non-native speakers: if you use it or anything like it, check the course policy first, disclose it, and keep your original draft. A pre-check with any detector, ours included, shows how one model reads your text, not what your school's tool will say.

5. Know the appeal routes

Find your institution's academic integrity policy and ask for the allegation and the evidence in writing. In England and Wales, students who finish their university's internal process can complain to the Office of the Independent Adjudicator, whose July 2025 guidance says students should get "all relevant evidence, including detection software reports", and that "the responsibility is on the provider to prove that the student has done what they are accused of doing, not on the student to disprove it" (OIA casework note; OIA announcement, 15 July 2025). Your international student office or students' union can help you through the process.

What institutions should take from this

The researchers on both sides agree more than the headlines suggest. Liang's team cautions against using detectors to assess non-native writers; Turnitin says its score "should not be used as the sole basis for action"; ETS's Hao calls for careful evaluation before consequential decisions; and the 2026 editing study says outputs are not definitive evidence. Vanderbilt cited the non-native bias finding when it switched Turnitin's AI detector off on 16 August 2023, and the University of Waterloo cited bias "toward students whose first language is not English" when it stopped using it as of September 2025.

Your writing is yours, however it sounds in English. A detector score is one model's guess about your wording, and the record of how you wrote is what settles the question.