Mandarin to English: Why Chinese ESL Writing Confuses AI Detectors

A student in Shanghai writes a careful, honest essay in English. She drafts it herself, checks her grammar twice, and turns it in. The AI detector lights up red. Nothing was generated, nothing was copied, and yet the tool insists a machine wrote it.

This happens far more to writers who learned English as a second language, especially native Mandarin speakers, than it does to native English writers. It isn’t bad luck, and it isn’t a sign your writing is weak. It’s a predictable result of how detectors actually work.

Key takeaways

• A 2023 Stanford study found AI detectors wrongly flagged a majority of TOEFL essays by non-native writers as AI-generated, while almost never misflagging native US student essays.
• The trigger is statistical, not a judgment about quality: detectors reward “surprising” wording, and careful second-language writing tends to be predictable.
• Mandarin-to-English habits, topic-first sentence order, repeated connectors, safe vocabulary, parallel clauses, read as “machine-like” to a detector.
• A flag is a probability, not proof. The same essay can score very differently across tools and across days.
• You can lower false positives by adding specific detail and varying your rhythm, not by writing worse English.

What the research actually found

In 2023, a Stanford team led by Weixin Liang published a study in the journal *Patterns* with a blunt title: “GPT detectors are biased against non-native English writers.” They ran real essays through seven popular detectors. On essays by native-English US eighth graders, the tools were close to flawless, an average false-positive rate of about 5%. On TOEFL essays written by non-native speakers, the same detectors averaged a 61% false-positive rate. More than half of the honest, human-written essays got tagged as machine output.

That’s not a small wobble. A tool people treat as an honesty check was getting it backwards most of the time for one specific group, the group that, in practice, includes a lot of Chinese students.

The researchers traced the cause to a single technical idea: perplexity. Low perplexity means the next word is easy to predict. The essays detectors flagged most confidently tended to have lower perplexity than the rest. Non-native writers stay inside standard, “safe” sentence patterns, which keeps perplexity low. Detectors read low perplexity as a fingerprint of a language model. So they punish exactly the caution a second-language writer was taught to have.

The same paper showed the bias running in reverse, too. When the team used ChatGPT to swap in fancier vocabulary on the non-native essays, the false-positive rate dropped sharply. When they simplified the native essays to sound more “ESL,” the flag rate shot up. The detectors weren’t reacting to who wrote the text. They were reacting to how predictable the words were.

Why Mandarin-to-English writing fits the “AI” profile

English coming out of a Mandarin-first mind carries habits from Chinese sentence structure, plus the patterns drilled in English class. Most of those habits are perfectly correct. They just happen to overlap with what AI text looks like statistically. A few of the big ones:

Topic-comment ordering

Mandarin often leads with the topic, then comments on it: “This problem, we can solve it from three aspects.” Translated into careful English, that becomes the very even, front-loaded sentence shape detectors associate with generated prose, “There are three main aspects to consider when addressing this problem.”

Connectors as scaffolding

English teachers in China lean hard on discourse markers, *firstly*, *moreover*, *in addition*, *in conclusion*, because they’re easy to grade and they signal structure. The trouble is that AI models lean on the same connectors. A paragraph that opens with “Firstly,” the next with “Moreover,” and the last with “In conclusion,” looks templated to a detector even when a real person wrote every word by hand.

Safe, frequent vocabulary

A second-language writer reaches for words they’re sure about. That’s smart, but high-frequency, expected words are precisely what keep perplexity low. A native speaker might write “the policy backfired”; a careful ESL writer writes “the policy did not achieve the expected result.” The second version is correct, clear, and statistically bland, which the tool reads as machine output.

Parallel structure everywhere

Mandarin rhetoric loves balance and symmetry. Carried into English, you get clean parallel clauses: “It improves efficiency, it reduces cost, and it increases satisfaction.” Nice writing. Also relentlessly even in rhythm, which is the opposite of the choppy, unpredictable cadence detectors read as “human.”

If you want the underlying math, we break down perplexity and burstiness, the two statistics behind most detectors in plain language. The short version: detectors don’t grade your ideas. They grade how surprising and how uneven your text is.

A mini-scenario: two sentences, same meaning

Say Wei wants to write about why his hometown built a new subway line. His careful draft:

> “The new subway line is very important. It can reduce traffic, it can save time, and it can protect the environment. Therefore, the city decided to build it.”

Every word is his. It’s also a near-perfect storm of low perplexity: predictable vocabulary, parallel clauses, a textbook connector, even sentence lengths. A detector will likely flag it.

Now the same point, written with specifics:

> “Chengdu’s third subway line opened in 2020, mostly because rush-hour traffic on the Second Ring Road had gotten unbearable, my dad’s commute used to take 50 minutes for nine kilometers. The line cut that to about 20.”

Same student, same honesty. But now there’s a real place, a year, a number, a personal detail, and the sentence lengths jump around. The writing got *better*, and the perplexity and burstiness went up as a side effect, which pulls the AI score down. That’s the move: you’re not tricking the tool, you’re writing more like a specific person and less like a template.

What this means if you’re getting flagged

First, a flag is not a verdict. Detectors output a probability, not proof. OpenAI quietly retired its own AI-text classifier on July 20, 2023, saying it had too “low a rate of accuracy” to keep up. Run your essay through three different tools and you’ll often get three different scores. Run it twice through the same tool after small edits and the number moves. Anyone treating that number as evidence of cheating is misusing it.

Second, protect yourself with process. If you’re a student, keep your drafts, your outline, your notes, your browser history of sources, the version history in Google Docs or Word. A false positive is much easier to push back on when you can show the essay grew over time instead of appearing fully formed.

Third, improve the writing rather than sabotage it. Don’t add random typos or strange words to fool the detector, that backfires with human readers and can look manipulated. Do these instead:

• Replace one vague claim per paragraph with a concrete fact, name, date, or number.
• Vary your sentence lengths on purpose. Put a four-word sentence next to a thirty-word one.
• Cut one templated connector per paragraph. You rarely need both “moreover” and “in addition.”
• Read it out loud. Where it sounds like a recited list, break the pattern.

Where a humanizer fits, and where it doesn’t

If the words and reasoning are genuinely yours and a detector keeps mislabeling them, smoothing out the formulaic patterns is editing, not cheating. That’s the honest use of a humanizer: take human writing that happens to score “AI” because of second-language habits, and nudge its rhythm and word choice toward how a specific person actually talks.

The line you shouldn’t cross is generating an essay and passing the machine’s thinking off as your own. A tool can fix how predictable your prose is. It can’t, and shouldn’t, supply the ideas. For ESL writers especially, the goal is to sound like *you*, not to chase a number.

If you want to dig deeper before you try anything, browse the rest of our writing on detection, or see the plans when you’re ready.

Frequently asked questions

Why does my English essay get flagged as AI even though I wrote every word myself? Most detectors don’t read for meaning. They estimate how predictable your words are (perplexity) and how much your sentence lengths vary (burstiness). Careful second-language writing tends to be predictable and evenly paced, the same pattern AI produces. The 2023 Stanford study documented this directly, with detectors misclassifying most non-native TOEFL essays.

Are AI detectors actually biased against Chinese and other ESL writers? The peer-reviewed evidence points that way. Liang and colleagues found detectors close to flawless on native US essays but wrong on the majority of non-native TOEFL essays. They linked it to low perplexity from “safe” phrasing. It’s a structural blind spot in the tools, not a comment on your ability.

Should I make my writing more “random” to avoid being flagged? No. Breaking grammar or inserting odd words can lower a score, but it makes your writing worse and can look manipulated. Add real specificity and vary sentence length instead. That raises perplexity naturally, as a result of clearer writing.

If a detector flags my essay, does that prove I cheated? No. A detector gives a probability, not evidence. OpenAI retired its own classifier in 2023 for low accuracy, and scores swing between tools and between runs. A flag is a reason to look closer, never a verdict. Keep your drafts and version history if you ever need to defend your work.

Does running my own writing through a humanizer count as cheating? If the words and ideas are yours, smoothing the formulaic patterns that trip detectors is editing. The line is passing off machine-generated thinking as your own. For human writing that’s being false-flagged, the goal is to sound like you.

—

If a detector has mislabeled your honest, hand-written English, you’re not imagining it, and you’re not alone. The fix isn’t to write worse. It’s to write more specifically, in your own rhythm, and to know what the tool is really measuring. When you’re ready to see exactly which lines are tripping the score, run it through PaperBleach.