Skip to content
humanise.ai Forum

Are AI detectors biased against non-native English speakers?

Staff answerWritten by Bronwyn Ashby · English as an additional language0 replies

Short answer

Research says some are. A 2023 Stanford study found that several AI detectors wrongly labelled many essays by non-native English writers as AI, while handling essays by US-born students almost perfectly. A score is not proof. If English is not your first language, keep your drafts, declare your tools and ask how detection is used.

Written by Bronwyn Ashby · Detection and policy editor ·
Citehumanise.ai Forum, "Are AI detectors biased against non-native English speakers?", https://humaniseai.ai/forum/read/are-ai-detectors-biased-against-non-native-english-speakers/, accessed 2026-11-20.
  1. Detection and policy editor ·

    What the study actually found

    The paper is "GPT detectors are biased against non-native English writers" by Liang and colleagues, available on arXiv. The authors tested several widely used GPT detectors on writing samples from native and non-native English writers. Their abstract says the detectors consistently misclassified the non-native samples as AI-generated, while native samples were identified accurately. They also caution against using such detectors in evaluative or educational settings, because they may penalise non-native speakers.

    Stanford HAI's write-up of the study gives the numbers:

    Finding Figure reported
    Detectors on essays by US-born eighth graders Described as "near-perfect"
    TOEFL essays by non-native writers classed as AI 61.22%
    TOEFL essays flagged by all seven detectors 18 of 91
    TOEFL essays flagged by at least one detector 89 of 91

    Those essays were written by people, not by AI. The study's senior author, James Zou, says the cause is how detectors work: they typically score using a measure called perplexity, which correlates with the sophistication of the writing. Non-native speakers tend to score lower on measures such as lexical richness, lexical diversity and syntactic complexity. In plain terms, careful, simple, correct English can look "predictable" to a detector, and predictable text is what these tools treat as machine-like.

    What this does and does not tell you

    Be fair to the evidence. It is a 2023 study of seven detectors, using TOEFL essays. The HAI article does not say that the tool your university uses was among them, so do not assume the same figures apply to your submission. Detectors have also been updated since. What the study does show is that a bias of this kind is possible and measurable, and that a flag on its own is weak evidence about who wrote something.

    Turnitin's own position, set out in its FAQ on AI writing detection, is that its indicator should not be the sole basis for action against a student. The thread on what a Turnitin AI writing score means explains the number in more detail.

    Practical steps if English is an additional language

    1. Write in one document with version history on. Word on OneDrive and Google Docs both record your edits. A visible timeline of your own drafting is the most useful evidence you can have.
    2. Keep your notes and plans. Handwritten outlines, annotated readings, early bullet points and emails to your tutor all show how the work developed.
    3. Declare grammar and translation tools. If you used a grammar checker, a translation tool or an AI assistant to check your English, say so in the way your university asks. The thread on declaring AI use covers this. An honest declaration protects you far better than silence.
    4. Ask how detection is used. Your module leader, your department or your students' union can tell you whether AI detection is used on your course, and what happens if a report flags something. Ask whether a score alone can trigger a meeting.
    5. Use language support, not just software. Many universities have an English language centre or writing service. A human tutor who has seen your drafts can also speak to your writing development.
    6. Know your rights if you are flagged. The thread on what happens next when Turnitin flags your essay walks through the usual process.

    What not to do

    Do not rewrite your own honest work to make it more "complex" in the hope that it scores better. You may make it worse, and you are changing your own voice to suit a tool. Do not rely on free online detectors either: different tools give different results, and none of them is the one your university uses. The paper itself notes that simple prompting can change detector outcomes, which is a further reason these scores are unreliable, but we are not going to suggest tricks. The honest route is a clear record of how you wrote the piece and what help you had.

    If you are a marker or manager reading this

    A high score on work by a multilingual writer is a reason to look at the drafting history and talk to the person, not a conclusion. The same applies to employers who screen applications or reports.

    If you post a follow-up, tell us your course level and country. Rules and appeal routes differ between the UK, Australia and New Zealand, and the best advice depends on yours.

    1
    ReplyLink

Plain text or simple Markdown. Links are nofollow. Your email is never shown.

humanise.ai Forum. Questions and answers on using AI writing tools well at university and work, voice and tone, and AI detector worries.