Seven GPT detectors flag 61% of TOEFL essays as AI while native writing sails through

GPT detectors are biased against non-native English writers

Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, James Zou

cs.CL, cs.AI, cs.HC, cs.LG

2023-04-06

Seven GPT detectors mislabel 61% of human TOEFL essays as AI, versus 5% on US 8th-grade essays; a native-style rewrite drops false positives to 12%.

What problem this solves

In early 2023, schools and journals plugged GPT detectors into the pipeline, worried about generated homework and generated abstracts. Most of those detectors keyed off a single statistic: low perplexity, meaning a language model finds the text easy to predict. Native writers jump around in vocabulary and syntax, so they are hard to predict. Non-native writers use a narrower lexicon and tidier syntax, so they are easier to predict. The detector's fingerprint for "sounds like a second-language learner" is the same fingerprint it uses for "sounds like GPT."

The Stanford group around James Zou ran seven off-the-shelf detectors on human writing from both groups. The question is not whether detectors work on average. It is who they fail on.

Method

Two human corpora. Ninety-one TOEFL essays dated 2020 or earlier, scraped from a Chinese education forum, written before ChatGPT existed. Eighty-eight US 8th-grade essays from the Hewlett Foundation ASAP automated scoring set as the native-ish control. They also kept 70 real college application essays and 145 CS224n project abstracts for the generate-then-rewrite bypass tests.

The seven detectors, queried on 15 March 2023: Originality.AI, Quil.org, Sapling, OpenAI's public detector, Crossplag, GPTZero, ZeroGPT. Perplexity was measured with open-source GPT-2 XL.

Two interventions. First, ChatGPT was asked to "Enhance the word choices to sound more like that of a native speaker" on the TOEFL essays, and the reverse prompt simplified 8th-grade essays "as if written by a non-native speaker." Second, ChatGPT-3.5 (14 March snapshot) wrote 31 Common App essays after filtering refusals, plus 145 fake CS224n abstracts, then a second-round self-edit: "Elevate the provided text by employing literary language" or "advanced technical language."

To check that low perplexity is more than a student-essay artifact, they scored 1,574 ICLR 2023 accepted papers whose deadline and rebuttal window both predated ChatGPT, grouping authors by whether their listed country is majority native-English, and controlling for review score.

Results

SettingMean false-positive / detection
Human TOEFL essays61.22% flagged as AI
Human US 8th-grade essaysnear-perfect, 5.19% false positives
TOEFL after native-style rewrite11.77%
8th-grade after non-native simplification56.65%
ChatGPT college essaysup to 100% detected
Same essays after literary rewriteup to 13%
ChatGPT scientific abstractsup to 68%
Same abstracts after technical rewriteup to 28%

Eighteen of 91 TOEFL essays (19.78%) were flagged by all seven detectors; those 18 had significantly lower perplexity (p=9.74e-5). At least one detector flagged 89 of 91 (97.80%). After the native-style rewrite, only one essay still drew a unanimous AI vote.

ICLR abstracts follow the same pattern: non-native-country authors have lower perplexity (p=0.035), and the gap survives controlling for review rating (p=0.033). Similar-quality papers, easier-to-predict abstracts.

The detector is eating "easy to guess." Non-native writers are easier to guess. Default GPT output is also easier to guess. Raise the diction one notch and both labels flip.

Why it matters

New York City public schools had just banned ChatGPT. Journals were demanding disclosure. Detectors were being wired into homework and peer review. This paper pins the collateral damage: not people who use AI, people whose English is less ornate. The ugly corollary is that the cheapest way for a non-native writer to stop being mislabeled is to ask GPT to dress the sentences up. The detector pushes people toward the tool it claims to catch.

The operational takeaway is blunt. Do not put GPTZero-class products on homework, hiring, or review gates when a large share of the population writes English as a second language. Perplexity as an AI fingerprint is a systematic bias on L2 writing, not a threshold you can tune away. One rewrite prompt also punches through the 2023 commercial detectors, so they were not catching determined GPT users either.

Limitations

The authors list three. Sample sizes are small: 91 TOEFL, 88 8th-grade. Most detectors sit on a GPT-2 backbone; whether GPT-3 or GPT-4 backbones repeat the bias is untested. DetectGPT's second-order log-probability method was too expensive to run at scale and was left out.

A few more gaps sit in the design. The TOEFL set comes from one Chinese forum, so "non-native" here mostly means Chinese-L1 exam writing, not European or South Asian English. Eighth-grade essays are a young control; the Discussion sometimes talks about college essays as the native reference, while the headline numbers are ASAP grade 8. Country of affiliation is a noisy proxy for native language on ICLR, and international labs get mis-binned. This is a July 2023 snapshot. OpenAI later retired its detector and GPTZero has been rebuilt. Treat the result as a diagnosis of that generation of perplexity detectors, not a verdict on every detector shipping in 2026.

Terms

Source

What people are saying

Related papers

All paper explainers