AI Detector Bias: Are International Students Unfairly Flagged?

plagiarism-checker-online.net Editorial Team  |  Updated October 3, 2026

AI detectors can wrongly flag human-written English by non-native writers. In a 2023 study, the average false-positive rate across seven detectors on 91 TOEFL essays was 61.3%. That historical result is not an error rate for every current tool. This article explains the evidence and how to respond if your own work is flagged.

What the Research Found

Liang and colleagues published GPT detectors are biased against non-native English writers in Patterns 4(7), article 100779 (2023), DOI 10.1016/j.patter.2023.100779. They tested seven GPT detectors on 91 TOEFL essays collected from a Chinese forum and 88 US eighth-grade essays from the Hewlett Foundation's ASAP dataset. They did not recruit essay writers from 91 countries.

The seven detectors incorrectly flagged the TOEFL essays at a mean rate of 61.3%, while classification of the US essays was near perfect. This comparison used different essay datasets, not a controlled experiment with matched college-level assignments. It demonstrates a risk in the tested 2023 systems, not a current performance estimate for Turnitin or every detector.

Do not apply this percentage to your own university or language background. A relevant evaluation would need to test the current detector version on writing comparable to your assignment.

Why Does This Bias Exist? The Technical Explanation

Some approaches use perplexity or variation in sentence structure, but vendors use different models and signals. Liang et al. linked false flags in their tests to low perplexity.

Perplexity is a measure of how "surprising" word choices are in context, one of the core signals explained in our guide to detecting AI-generated text. Language models like ChatGPT generate text by selecting statistically probable words given the preceding context. This produces low-perplexity text, text where each word choice is unsurprising given what came before. Human creative writing is typically higher-perplexity because individuals introduce unexpected vocabulary, idiosyncratic constructions and personal voice.

Burstiness refers to variation in sentence complexity. Human writing tends to mix long, complex sentences with short, simple ones. AI-generated academic text tends to maintain more uniform sentence length and structure throughout a document.

Here is the problem: non-native English speakers who have learned academic English through formal instruction tend to write low-perplexity, low-burstiness text, not because they used AI, but because the academic English they learned is itself modeled on predictable, grammatically correct formal registers. The rules of academic writing they absorbed (clear topic sentences, appropriate vocabulary, standard sentence structures) produce writing that is statistically similar to AI output. Their writing does not sound like casual native English because it is not casual native English: it is careful, rule-following formal prose that happens to share statistical properties with AI-generated text.

Specific Language Backgrounds and Risk Levels

The study's TOEFL sample came from a Chinese forum. It does not establish separate false-positive rates for Chinese, Korean, Japanese, Arabic, or other first-language groups.

The authors linked false flags to limited linguistic variability and low perplexity. That is a warning against treating predictable wording as proof of AI use, not a way to rank individual students by nationality or fluency.

Institutional Responses: How Universities Are (and Are Not) Adapting

Institutional approaches differ, as our guide to university AI policies explains. Check your own institution's current procedure. Turnitin's guidance says its AI report should not be the sole basis for adverse action against a student.

However, implementation is uneven. Many institutions continue to use detection scores as a primary trigger for investigation with minimal acknowledgment of the false positive issue. Students, particularly international students who may be unfamiliar with their institution's processes and less confident challenging authority, are disproportionately vulnerable.

Tool providers have acknowledged the problem. Turnitin in particular has been explicit that its AI detection scores should be treated as indicators rather than determinations, and has recommended against using the technology as the sole basis for academic misconduct allegations. Whether this guidance is being followed in practice varies substantially by institution.

What You Should Do as an International Student

Before Submission: Document Your Process

The most effective protection against a false positive is evidence of your writing process. Save every draft of your paper with timestamps. Keep your research notes, source lists and outlines. Note the dates and times you worked on the paper. Many word processors and cloud storage platforms (Google Docs, Microsoft OneDrive) automatically version-track documents; make sure this is enabled.

Check your institution's specific position on AI use; our guide to AI writing in academic papers maps what is typically allowed and what is forbidden. Before you submit, also consider running your paper through an AI checker yourself. This gives you a pre-submission view of how your paper scores. If it returns an unexpectedly high score on clearly human-written work, you are forewarned; you can prepare documentation and, if appropriate, raise the issue proactively with your instructor before submission rather than reacting defensively afterwards.

If You Are Flagged: The Appeal Process

If your paper is flagged with a high AI score and you are accused of using AI improperly, the following steps are important:

• Request a copy of the detection report. Check your university's procedure for requesting the evidence on which the allegation is based. The report should show which specific passages were flagged and with what reported indicators.
• Gather your process documentation. Bring draft versions, notes, source materials and any other evidence of your writing process. The existence of iterative drafts over time is strong evidence of human authorship. Our guide on what to do if you are falsely accused of using AI has an evidence checklist that shows where to find each item.
• Request an oral examination. Ask to discuss your paper in a face-to-face meeting or oral examination. If you wrote the paper yourself, you should be able to discuss its argument, explain your choices and answer questions about the content, something AI cannot do on your behalf. Offering to do this proactively signals confidence in your authorship.
• Cite the research literature if necessary. Refer to published research on false positives for non-native speakers. The Liang et al. (2023) study in Patterns is a peer-reviewed source you can cite in any formal appeal. Your university's equity and inclusion office may also be a resource.
• Contact student support services. Most universities have services specifically for international students, as well as academic integrity advisors and student union representatives who can support you through a formal process. Do not navigate a misconduct proceeding alone.

A Systemic Problem Requiring Systemic Solutions

The study shows why detector results need scrutiny. It does not establish that future models cannot improve fairness or that watermarking will solve the problem. Our guide to AI watermarking and SynthID explains the limits of provenance signals.

Institutions that use AI detection responsibly acknowledge this and build their processes accordingly. Those that treat detection scores as definitive are not only applying an unreliable tool incorrectly; they are at risk of systematically disadvantaging students who are already navigating substantial barriers in higher education. Awareness of this issue, and advocacy for fair process, is important for students and educators alike. Students can also take proactive steps to produce clearly original work; our guide on how to avoid plagiarism covers the foundational practices that support genuine academic authorship.

Sources checked October 3, 2026: Liang et al., Patterns (2023); Turnitin report guidance.

Frequently Asked Questions

How strong is the evidence that AI detectors flag non-native English writers?

The best-known study is Liang et al. (2023) in Patterns. They ran seven GPT detectors over 91 TOEFL essays by non-native speakers and over essays by US eighth graders. The TOEFL essays were wrongly flagged as AI-generated at an average rate of 61.3%, while the US students' essays were classified almost perfectly. The detectors were 2023 versions and the sample was modest, so current tools may differ, but the authors cautioned against using detectors in educational settings, especially for non-native speakers.

Should I write less formally to avoid being flagged?

No. Deliberately weakening your writing, or running it through a humanizer tool, lowers its quality and can look like concealment if your work is questioned. Evidence of your process protects you better than a style change. If you want to know how your paper scores, check it before submitting and raise the issue with your instructor if a high score shows up on your own work.

What evidence shows I wrote the paper myself?

Version history is the strongest. Google Docs and Word files saved on OneDrive keep timestamped versions, so turn that on before you start writing. Add your outline, reading notes, saved source PDFs and earlier drafts. If you're asked, be ready to explain your argument and your source choices in a short conversation.

What should I do first if I'm accused of using AI?

Ask for the detection report and the allegation in writing, and don't delete or rewrite anything. Gather your drafts and version history, then check your university's misconduct procedure for how and when to respond. Student advisory services or the students' union can usually come with you to the meeting.

This article is part of our AI Detection Guide.