AI Detectors Are More Likely to Flag Non-Native English Writers. Here’s the Research.
· El Mehdi

If you’ve ever had your writing flagged as "AI-generated" and you learned English as a second language, there’s real research explaining why that might not be your fault.
The study: a 61% false-positive rate
In 2023, a team of Stanford researchers led by Weixin Liang tested seven popular AI-detection tools — Originality.AI, Quill.org, Sapling, OpenAI’s own classifier, Crossplag, GPTZero, and ZeroGPT — against 91 essays written by non-native English speakers for the TOEFL exam. Every one of these essays was written by a person.
The detectors flagged 61.22% of them, on average, as AI-generated. Nearly 1 in 5 (19.78%) were flagged as AI by all seven detectors at once. By contrast, the researchers report that essays written by native-English-speaking US eighth-graders were classified accurately. The TOEFL sample is small (91 essays), so the exact percentage should be read as an indication of a pattern rather than a precise rate.
The paper, "GPT detectors are biased against non-native English writers", was later published in Cell Press’s Patterns journal.
Why this happens
Many AI detectors rely on measures such as "perplexity" — roughly, how predictable a text’s word choices are. Text generated by large language models tends to be low-perplexity: smooth, grammatically safe, and statistically unsurprising.
The problem is that non-native speakers — especially people trained with standardized test prep, like TOEFL study guides — also tend to write in simpler, more formulaic, more "predictable" English. Not because they used AI, but because that’s how they were taught to write safely in a second language. The researchers found the detectors could not reliably tell the difference.
The same study found that simple prompting — asking a model to rewrite its output in more elaborate language — let AI text slip past the detectors, so the people most likely to be wrongly flagged are often not the ones the tools are meant to catch.
Universities are noticing
This finding has influenced real policy. In August 2023, Vanderbilt University disabled Turnitin’s AI detector. Its announcement noted that Turnitin had claimed a 1% false-positive rate at launch, which could have meant around 750 incorrectly flagged student papers, and concluded: "we do not believe that AI detection software is an effective tool that should be used."
The University of Waterloo also discontinued Turnitin’s AI-detection feature as of September 2025; its announcement cites the tool’s expense among its reasons.
If you’ve been flagged
A percentage score from a detector is not proof of anything. A few practical steps if this happens to you:
- Ask specifically what the score was based on. Many tools — and many instructors — can’t explain which parts of your writing triggered the flag.
- Point to the research above if it’s relevant to your situation, especially if English is your second language.
- Where possible, show your work: drafts, notes, revision history — anything that documents how you actually wrote the piece.
That is why WriteProof works differently. Our free check doesn’t just hand you a verdict — it shows the specific sentences and patterns that influenced the score, so you can see and respond to what’s being questioned.
Our optional Writing Timeline and Session Record let you keep a record of how a draft developed as you write it. They add useful context, but they cannot prove authorship on their own: browser-reported activity is not independently verified.
Sources
- Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). "GPT detectors are biased against non-native English writers." Patterns, Cell Press.
- Vanderbilt University, Brightspace (Aug 16, 2023). "Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector."
- University of Waterloo (2025). "Discontinuing use of AI detection functionality in Turnitin — September 2025."
Leave a Reply