A 2023 study in Patterns tested seven GPT detectors on 91 TOEFL essays written by non-native English writers; the average false-positive rate was 61.22%. That result describes those tools and essays in 2023, not every writer or today's version of every detector, but it is a strong reason not to treat a score as proof. HumaText can propose a phrasing edit on writing you own, with every change left for your review.

What the study found

The study is Liang, Yuksekgonul, Mao, Wu, and Zou, "GPT detectors are biased against non-native English writers", published in Patterns in July 2023. The researchers ran seven widely used GPT detectors on 91 human-written TOEFL essays from a Chinese forum and 88 essays by U.S. eighth-graders. On the TOEFL essays, the average false-positive rate was 61.22%. All seven detectors flagged 18 of the 91 essays (19.78%), and at least one detector flagged 89 of them (97.80%). The U.S. essays were classified with near-perfect accuracy.

Those findings apply to the tested samples, tools, and 2023 versions. They do not establish a current false-positive rate for every non-native writer or every detector. They do show that language background and writing style can confound automated authorship estimates.

Why it happened, according to the authors

The authors point to perplexity, a measure of how predictable a text's word choices are to a language model. Detectors that rely on it tend to treat low-perplexity text as more likely to be machine-generated. Writers working in a second language often rely on a smaller set of words and constructions they are confident in, which can make their prose more predictable. In the study, the TOEFL essays that every detector flagged had significantly lower perplexity than the others.

Commercial detectors are proprietary and do not all use the same inputs, so the mechanism may differ by tool. The broader lesson holds: a style-based score describes how text reads, not who wrote it.

If your work is flagged

Ask for the institution's process and what evidence it accepts. Preserve your drafts, notes, sources, and version history, and be ready to explain how the document developed. A detector score should not be treated as standalone proof, and the study above is one reason why.

Polishing a paper is not a way to dispute or change a detector result, and HumaText does not promise any detector outcome. Its own AI-pattern score is a rough writing-style heuristic, not proof of authorship and not a forecast of any other tool.

What a HumaText pass can do for ESL writers

Plainer constructions. Academic English written in a second language often leans on formal constructions that read stiffly. The model is prompted to replace inflated words with plain ones of the same strength and to cut filler: "in order to" becomes "to," "due to the fact that" becomes "because," "make use of" becomes "use."

Varied sentence length. A paragraph of four similar-length sentences can feel monotonous. The model is prompted to vary length where the content allows.

No slang. Standard and Aggressive modes are instructed not to introduce slang, idioms, or figurative language, which helps if you want the result to stay clear and formal.

Meaning first. The model is told to keep every claim, hedge, and condition and to copy names, numbers, and citations exactly. No automated rewrite can guarantee that, and meaning shifts can be harder to spot in a second language. Compare the source and output carefully, especially around claims and technical terms.

Illustrative example

Draft: "It is important to note that the methodology utilized in this study was designed for the purpose of providing comprehensive data regarding the influence of social media upon consumer behavior."

A Standard pass might return something like: "This study's methodology was designed to provide thorough data on how social media influences consumer behavior."

The "it is important to note" preamble is gone, "utilized" was dropped because "designed" already carries the meaning, and "regarding the influence of X upon Y" became "how X influences Y." The claim is the same. Actual output varies from run to run.

Keep a record of your process

Independent of any polish tool, keep version history, notes, sources, and earlier drafts. This record can show how the document developed and gives a reviewer better evidence than a score alone.

If you are flagged, follow the institution's process and ask what evidence it accepts. Version history supports your account, but it is not automatically conclusive.

Use it where your course permits

Check your course or institution's policy on AI-assisted editing before using any tool on submitted work, and disclose tool use when the policy requires it.

Free to try

You can try HumaText at humatext.com without an account; the anonymous allowance is 250 words per rolling 24 hours. Read every rewrite and keep only the changes that accurately express your own meaning.

ESL writing polish for non-native English writers

Key points