Why ESL and Non-Native English Writers Get Flagged by AI Detectors

Picture a student who studied English for years, sat the TOEFL, moved across the world to attend university, and writes every essay by hand without any help from a chatbot. Now picture an AI detector telling her professor that her paper is probably machine-generated. That isn’t hypothetical. It’s one of the most reliably reproducible failures in the whole AI-detection space, and the people it hurts most are the ones doing nothing wrong.

If you write English as a second (or third, or fourth) language, you deserve to know why this happens and what you can actually do about it.

Key takeaways

• A 2023 Stanford study ran 91 non-native TOEFL essays and 88 native US eighth-grade essays through seven popular GPT detectors. On average the detectors wrongly flagged the TOEFL essays as AI 61.22% of the time, while staying near-perfect on the native-speaker essays.
• The cause is mostly mechanical. Many detectors lean on *perplexity*, and clear, careful ESL writing tends to use common, predictable words that read as “low-perplexity” to a model.
• The strangest finding: enriching those same essays with fancier vocabulary using ChatGPT dropped the false-AI rate from 61.22% to 11.77%. More AI made the writing look *more* human to the tools.
• A detector score is a probability the vendor publishes, not proof. A flag is a reason to ask questions.
• The honest fix is to write in your own real voice and keep a record of your process. Counterintuitively, a rougher, more personal draft often reads as more human.

The study that put a number on it

In 2023, a Stanford research team led by Weixin Liang published a paper in the journal *Patterns* with a blunt title: “GPT Detectors Are Biased against Non-Native English Writers.” They gathered two sets of essays. One set: 91 TOEFL essays written by non-native English speakers. The other: 88 essays by US eighth-graders, all native speakers. Then they ran both through seven widely used GPT detectors.

The detectors were near-perfect on the native-speaker writing. On the non-native writing they were wrong more than half the time. Averaged across the seven tools, 61.22% of the TOEFL essays were misclassified as AI-generated (Liang et al., 2023). And it skewed worse at the extremes: all seven detectors *unanimously* tagged 18 of the 91 essays as AI, and at least one detector flagged 89 of the 91. These were human-written essays, by humans, with zero AI involvement.

Two sets of essays, the same task, the same tools. The only thing that changed was whether English was the writer’s first language, and the outcome flipped from near-perfect to coin-flip-or-worse.

So why does this keep happening?

Short version: a lot of detectors are looking at the wrong thing, and the wrong thing happens to correlate with being an ESL writer.

Most statistical detectors revolve around a metric called perplexity. GPTZero, for example, describes its own approach as estimating “how likely an AI model would have chosen the exact same set of words as found in the document.” Low perplexity means the words are predictable, the kind a language model would probably pick. High perplexity means the text keeps surprising the model with unusual choices.

Here’s the problem. AI models generate text by repeatedly choosing safe, high-probability words. That produces low perplexity. But you know who *else* tends to reach for common, dependable, high-frequency words? Someone writing carefully in a language they learned later in life.

It’s not a flaw in the writer. It’s a feature of how people write in a second language. You pick words you’re confident are correct. You avoid slang you’re not sure about. You keep sentence structures clean because clarity matters more than flourish. All of that is good, careful writing, and it produces exactly the low-perplexity fingerprint that perplexity-based detectors read as “machine.”

The Stanford authors said as much. They pointed to low perplexity in non-native writing, driven by less linguistic variability and richness, as the likely culprit. A native speaker tosses in an idiom, a weird verb, a regional turn of phrase, and the perplexity spikes. The detector reads that spike as humanity. The careful ESL writer, who left the idioms out on purpose, gets penalized for it.

The result that flips the whole thing on its head

Then the researchers did something clever, and what they found should bother anyone who trusts these tools.

They took the flagged TOEFL essays and prompted ChatGPT to rewrite them with more sophisticated, native-sounding vocabulary. The average false-AI rate fell from 61.22% to 11.77% (Liang et al., 2023). The share of essays *every* detector unanimously called AI dropped from 18 essays to a single one.

Read that again. Adding *more* AI involvement made the essays look *more* human to the detectors. The tools rewarded fancier word choice, not actual authorship. A genuinely human essay got flagged; a model-polished one sailed through. That isn’t a small calibration issue. The detector is measuring vocabulary sophistication and calling it humanity, and the two have almost nothing to do with each other. If you write your own honest prose in plain second-language English, you lose. If you launder it through a chatbot first, you win. That’s backwards, and it’s the clearest single piece of evidence that the signal these tools key on is the wrong one.

A quick scenario: meet Mariana

Let’s make it concrete. (The numbers below are illustrative, not from the study, but the pattern is exactly what the research describes.)

Mariana is a second-year nursing student. Portuguese is her first language. She writes a four-page reflection on a clinical placement, entirely herself, over three evenings. She’s proud of it. Her prose is plain and direct because that’s how she thinks in English: “The patient was anxious. I checked his chart. I asked the charge nurse for guidance before I gave the medication.”

Short sentences. Common words. Clear structure. No idioms, no showing off.

Her instructor runs it through a detector. It comes back 88% likely AI.

Nothing in Mariana’s writing is wrong. The very qualities that make her essay readable, which is supposedly what an instructor wants, are the qualities the detector reads as robotic. A native-speaker classmate who wrote a messier essay full of casual asides and odd phrasing scores 4% AI. He didn’t write a better paper. He wrote a *burstier* one.

So a score in isolation tells you very little. The number can’t see Mariana’s three evenings of work, her drafts, or her honesty. It only sees perplexity. And that’s exactly the situation where students can and should defend themselves with evidence rather than apologies.

The detectors themselves admit they’re not certain

It’s worth remembering who tends to be most skeptical of AI detectors: the people who build them.

OpenAI launched its own AI text classifier in early 2023 and quietly shut it down that July, citing a low rate of accuracy. Its own numbers were rough. The tool correctly identified only about 26% of AI-written text while falsely flagging human writing 9% of the time. The company that arguably understands these models best looked at its own detector and pulled the plug.

Mainstream detectors today output a probability or a likelihood, not a verdict. That word matters. A 90% score is a statistical guess from a tool that, by its own vendor’s framing, can be confidently wrong. We’ve written separately about why detectors will never hit 100% accuracy, and ESL bias is one of the sharpest illustrations of the limit. Treat any score as a starting point for a conversation, never the end of one.

What ESL writers can actually do

You can’t rewrite the detectors. But you have more control than you think.

Keep a paper trail, and respond with evidence if you’re flagged. This is the single most powerful protection, and it’s also where this post hands off to our dedicated guide. The short version: write in Google Docs or Word so your version history is captured automatically, then, if a flag comes up, ask which tool was used and what it scored before you say anything else. A growing document with edits and dead ends is hard to argue with. For the full step-by-step on handling a wrongful accusation, see how students can defend themselves. It’s also worth knowing the broader reasons detectors throw false positives so you can explain the pattern, not just point at your timeline.

Write in your own voice, not a “smarter” one. There’s a temptation, after reading about this bias, to inflate your vocabulary to dodge the perplexity trap. The study even shows it works on the tools. But faking a fancier voice carries real costs: you risk burying your own meaning, and an instructor who knows your writing may notice the sudden shift. Vary your sentence lengths the way you naturally would in conversation. Mix a short sentence with a long one. Let a little of *you* show up on the page, idioms and all.

Add specifics only you would know. Concrete, personal detail is the enemy of generic, low-perplexity text. The name of the street where your placement was. The exact thing your professor said in week three. A small mistake you made and what you learned from it. Models produce generic filler because they don’t have your Tuesday. The more your writing carries details no one could have invented for you, the less it resembles the safe, averaged-out prose a model generates.

Use a checker to understand your own writing, not to game anyone. Running your own draft through a detector before you submit is about literacy. If a tool flags a paragraph you wrote yourself, that’s useful information: it tells you which parts of your natural style happen to trip the perplexity wire, so you’re not blindsided in front of a professor. You can run your draft through a free checker and see which sentences light up, then decide for yourself whether to adjust them or simply keep your receipts. Just keep in mind the same caveat that runs through this whole piece: detectors are documented to misfire on human text, so read the result as a signal, not a sentence.

Where this leaves us

The ESL bias in AI detectors isn’t a glitch that gets patched next quarter. It’s baked into the math of how perplexity-based detection works, and it lands hardest on exactly the writers least likely to be cheating. The rewrite experiment proves the point in a single, uncomfortable line: dressing human writing up with AI made it *less* suspicious, because the tools were never really measuring authorship to begin with.

Until institutions internalize that a probability isn’t proof, the burden unfairly shifts to writers to protect themselves. So protect yourself. Keep your drafts. Write like you, not like a thesaurus. And when a tool reduces three evenings of honest work to a single suspicious number, remember what that number actually reflects: how predictable your word choices were, and nothing about whether you wrote them.

Try it yourself

You don’t have to take a detector’s word for it. Paste your draft into PaperBleach’s free AI checker and humanizer to see how it scores — and make it read more like you actually wrote it.