The Stanford Study That Proved AI Detectors Are Biased Against Non-Native English Writers

If you write in English as a second language and an AI detector has ever flagged your work, here’s something worth knowing: a team at Stanford ran the numbers, and the tools weren’t catching cheaters. They were catching you for sounding like you. The research is real, it’s peer-reviewed, and the findings are blunt.

Key takeaways

• A 2023 Stanford study (Liang et al., published in *Patterns*) tested 7 popular GPT detectors and found they misclassified 61.22% of human-written TOEFL essays as AI-generated, while scoring near-perfectly on essays by native English speakers.
• The cause is perplexity. Detectors read predictable, low-variability word choice as a sign of AI, and non-native writers naturally use a smaller, safer vocabulary.
• All 7 detectors agreed on flagging 18 of 91 TOEFL essays; 89 of 91 were flagged by at least one. That’s a false-positive problem, not a cheating problem.
• When researchers asked ChatGPT to enrich the vocabulary of those same essays, the false-positive rate dropped to 11.77% – evidence the tools react to style, not authorship.
• The authors openly cautioned against using AI detectors in schools, because the penalty lands hardest on people writing in a second language.

What the study actually tested

In 2023, Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou at Stanford set up a clean experiment. They gathered 91 essays written by non-native English speakers for the TOEFL exam, and 88 essays written by US eighth-graders, native speakers. Every one of these was human-written. No AI involved. Then they fed all of them to seven widely used GPT detectors.

The detectors handled the eighth-grade essays fine. Near-perfect accuracy, correctly labeling them as human. The TOEFL essays were a different story. On average, the detectors called 61.22% of them “AI-generated.”

Read that again. More than six in ten essays, all written by real students sitting a real exam, got tagged as machine output. And it wasn’t one rogue tool dragging down the average. All seven detectors unanimously flagged 18 of the 91 essays. Stretch it to “flagged by at least one detector” and you hit 89 of 91 – basically everyone.

Why the detectors got it so wrong

The short answer is a metric called perplexity. The longer answer is worth your time, because it explains exactly why this bias is baked in rather than accidental.

Perplexity measures how surprising each word is, given the words around it. If text is highly predictable – common words, expected phrasings, simple structure – it has low perplexity. If it zigzags with rare words and unusual turns, perplexity is high. Many detectors assume AI writes low-perplexity text (which it often does) and humans write high-perplexity text. So low perplexity gets read as “probably a machine.”

Here’s the trap. Non-native English writers also tend to produce low-perplexity text. Not because they’re worse writers, but because second-language writing leans on a safer, more limited vocabulary and steadier sentence patterns. The Stanford team confirmed this directly: the TOEFL essays that all seven detectors flagged had significantly lower perplexity than the ones that slipped through (p = 9.74E-05). The detector wasn’t detecting AI. It was detecting linguistic caution and mistaking it for a machine.

If you want the full mechanics of how this scoring works, we broke it down in how perplexity and burstiness drive these tools.

The experiment that proved it was style, not authorship

This is the part that turns a worrying finding into a damning one. The researchers took the same flagged TOEFL essays and gave ChatGPT a simple instruction: “Enhance the word choices to sound more like that of a native speaker.”

Same ideas. Same students’ work. Just richer vocabulary. The average false-positive rate fell from 61.22% to 11.77%, and the number of essays unanimously flagged as AI dropped from 18 to a single one.

Then they ran it in reverse. They took the native-speaker eighth-grade essays and asked ChatGPT to “simplify word choices as if written by a non-native speaker.” The misclassification rate climbed from 5.19% to 56.65%.

So the same detector, looking at work by the same human authors, would flip its verdict based purely on how fancy the vocabulary was. That’s not authorship detection. That’s vocabulary policing with a confidence score attached.

A quick scenario to make it concrete

Picture two students submitting the same argument in a composition class.

Maria grew up speaking Spanish. She writes: “The author makes a good point. He says that money does not buy happiness. I agree with this idea because rich people are also sad sometimes.”

Her classmate Jake, a native speaker, writes: “The author’s claim – that wealth fails to guarantee contentment – resonates, particularly given how often affluence coexists with quiet despair.”

Same point. Jake’s sentence is full of less-predictable word choices, so a perplexity-based detector reads it as comfortably human. Maria’s is clean, direct, and predictable, so the same detector may flag it as AI. Neither used AI. One of them just learned English second. The Stanford data says Maria is the one who gets called in for an academic-integrity meeting.

Why this matters beyond a single paper

The study’s authors didn’t bury the implication. They wrote that institutions should be cautious about using these detectors in evaluative or educational settings, precisely because the false positives concentrate on non-native speakers. When a tool is wrong in a random way, that’s noise. When it’s wrong in a way that consistently targets one group, that’s bias with consequences – failed assignments, integrity hearings, and a chilling effect on students who already face enough friction.

There’s also a bitter irony the paper points out. To dodge a false flag, a non-native writer might feel pushed to run their own honest work through an AI tool just to “sound native” enough to pass. The detector, meant to police AI use, ends up nudging innocent people toward AI. That’s the system running backwards.

A few honest caveats, because the study spells them out too. The sample sizes were small – 91 TOEFL essays and 88 eighth-grade essays. Most of the detectors tested used older GPT-2-era models as their backbone, which the authors named as a limitation. They framed the whole thing as a pilot study, not the final word. But the mechanism – perplexity-based detection penalizing predictable writing – is still alive in plenty of tools today, which is why this paper keeps getting cited.

What this means if you’re the one getting flagged

If you write in a second language and you’re staring at an “85% AI” score on something you wrote yourself, the study offers a strangely empowering reframe: the tool isn’t reading your intent or your honesty. It’s reading your sentence variety and word choice. That’s it.

So the useful move isn’t to panic or to confess to something you didn’t do. It’s to understand what the detector measures and to write with more variation – mix short sentences with long ones, reach for a more specific word now and then, break up repetitive structure. If you want to compare what different tools cost and offer before you rely on any of them, our pricing page lays it out, and you’ll find more guides on AI detection and writing covering the same ground in plain language.

Frequently asked questions

Who ran the Stanford study and where was it published? It was led by Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou at Stanford University. The paper, “GPT detectors are biased against non-native English writers,” appeared in the journal *Patterns* in July 2023. A preprint lives on arXiv (2304.02819).

What exactly did the study find? Across 7 widely used GPT detectors, the average false-positive rate on 91 human-written TOEFL essays was 61.22%, versus near-perfect accuracy on 88 US eighth-grade essays by native speakers. All 7 detectors unanimously flagged 18 essays as AI; 89 of the 91 were flagged by at least one.

Why do AI detectors flag non-native English writing? Most rely on perplexity, a measure of how predictable each word is. Non-native writers tend to use a narrower, safer vocabulary and simpler patterns, which produces low-perplexity text. AI also produces low-perplexity text, so the detector confuses the two. It’s measuring style, not who wrote it.

Does this mean ESL students should use an AI humanizer to dodge false flags? The honest answer is subtler. The study showed enriching vocabulary cut false positives sharply, which tells you these tools react to surface style. If you’re a non-native writer getting wrongly flagged, the goal is to vary your sentences and word choice so the statistics don’t misread your own work – not to disguise authorship of something you already wrote.

Is the study still relevant given how much detectors have changed? The exact numbers are from 2023, and most detectors tested used GPT-2-era backbones, which the authors flagged as a limitation. But the underlying mechanism – perplexity-based detection penalizing predictable writing – still powers many tools, so the bias warning remains one of the most cited cautions against treating detector output as proof.

The bottom line

A detector that flags 61% of honest essays from one group of writers isn’t catching cheating – it’s reproducing a bias, and Stanford put hard numbers on it. If you write in a second language, knowing what these tools actually measure is your best defense. Want to see how your own writing scores and where it gets misread before anyone else does? Run your draft through a detector and see the score.