Why ESL Students Get False AI Flags 2x More Often Than Native Speakers (And What the Research Says)

Imagine writing every word of an essay yourself, in a language you worked hard to learn, and then watching software label it “98% AI-generated.” For a lot of multilingual students, that isn’t a hypothetical. It happens, again and again, and the research explains why.

The short version: the ESL students false AI detection rate is dramatically higher than it is for native speakers, and it has nothing to do with whether you actually used AI. It has to do with how detectors measure writing.

Key takeaways

• A 2023 Stanford study (Liang et al.) tested seven AI detectors and found they flagged an average of about 61% of non-native English (TOEFL) essays as AI-written, while scoring native US student essays with near-perfect accuracy.
• The cause isn’t dishonesty. It’s *perplexity*: detectors flag predictable, limited-vocabulary writing, and that pattern shows up naturally in second-language English.
• In the same study, prompting a model to re-edit those essays with more varied, native-like word choices dropped the false positive rate from about 61% to roughly 12%.
• Detectors output a probability, not proof. A “likely AI” score is a guess, and that guess is systematically less fair to ESL writers.
• If your honest work gets flagged, you can document your process and ask for human review instead of accepting a score as a verdict.

What the research actually found

In 2023, a team at Stanford led by Weixin Liang published a study with a blunt title: *GPT detectors are biased against non-native English writers.* They ran seven widely used AI detectors against two sets of writing. One set was 91 TOEFL essays written by non-native English speakers, which the students had uploaded to a Chinese test-prep forum before ChatGPT even existed. The other was 88 essays by US eighth-graders.

The results split hard along language lines. On the native US essays, the detectors were nearly perfect, correctly recognizing them as human. On the TOEFL essays, they fell apart. The average false positive rate was 61.22%. More than half of genuinely human, non-native essays got tagged as machine output.

It gets worse when you look at how the tools agreed. About one in five of the TOEFL essays — roughly 19% — were unanimously flagged as AI by all seven detectors at once. And 89 of the 91 essays were flagged by at least one detector. So if a school runs an essay through several checkers and trusts any single “AI” verdict, a non-native writer is almost certain to trip at least one of them.

Compare that to the native essays, where the false positive rate sat around 5%. That’s not a “2x” gap. It’s far wider than that. The headline “twice as often” is honestly the *conservative* read of this data.

So why do detectors get it so wrong?

Here’s the part that surprises people: the detector isn’t reading your essay for meaning. It’s doing statistics.

Most AI detectors lean on a measure called perplexity — basically, how surprised a language model is by your next word. Write something unpredictable, full of unusual phrasing and varied vocabulary, and perplexity goes up; the text “looks human.” Write in steady, simple, predictable sentences, and perplexity drops; the text starts to “look like” a model, because models also produce statistically likely word sequences. (We unpack how perplexity and burstiness drive these scores in a separate piece if you want the full mechanism.)

Now think about how writing in a second language usually works. You lean on the vocabulary you trust, keep your sentence structures clean to avoid mistakes, and skip the risky idioms you’re not sure about. None of that is wrong — it’s careful, sensible writing. But it produces exactly the low-perplexity, predictable pattern that detectors are trained to suspect.

The Stanford team showed this was the mechanism, not a coincidence. When they prompted a model to re-edit the TOEFL essays “to sound more like that of a native speaker,” swapping in richer word choices, the measured perplexity went up and the false positive rate dropped from about 61% to roughly 12%. The detector was reacting to vocabulary complexity the whole time. Change the surface texture and you change the verdict, even though the underlying ideas stay the same.

A mini-scenario: Mariana’s discussion post

Picture Mariana, a grad student whose first language is Portuguese. She writes a 300-word discussion post about supply-chain risk. She knows the material cold, but she sticks to vocabulary she trusts: “important,” “shows,” “the company should,” clear and direct. No AI involved. She wrote every line at her kitchen table.

She pastes it into the course portal. The instructor’s tool returns “87% likely AI.” Mariana panics, because the score *feels* like an accusation, even though it’s really just a probability estimate of word predictability.

What actually happened: her clean, confident, low-perplexity prose hit the same statistical signature the detector associates with generated text. A native speaker who wrote the identical idea with a couple of offhand idioms and a long, rambling sentence might have sailed through. The content was equally human. Only the surface texture differed — and the texture is all the detector can see.

What ESL writers can reasonably do about it

None of this means you should manufacture fake quirks or stuff your essay with thesaurus words you don’t understand. That usually reads worse, not better. But a few honest habits genuinely help, and they line up with what the research showed.

Vary your sentence length and rhythm

Detectors react to uniformity. Mix a short, punchy sentence with a longer one that carries a couple of clauses. This is normal good writing, and it raises the “burstiness” detectors associate with human authors. It’s a craft improvement either way.

Use the words you actually know — just more of them

You don’t need rare vocabulary. You need a *bit more range* than your safest default. Swap one repeated “important” for “central” or “the part that matters.” The Stanford prompt that fixed most false flags simply asked for more native-like word choices, not for showing off.

Keep your receipts

Write in a tool that saves version history (Google Docs revision history, for instance). Keep your outlines and notes. If a score ever gets questioned, your draft trail is far stronger evidence of authorship than any detector score is of guilt.

Read your draft out loud

If it sounds like you when you say it, that’s a good sign. And if you’re curious what a check costs before you lean on one, you can see what a check actually costs so there are no surprises. Knowing where you stand beats finding out after the fact.

A note on what these scores mean

It’s worth repeating, because schools sometimes forget it: a detector outputs a probability, not a fact. OpenAI itself retired its own AI Text Classifier in July 2023, citing low accuracy. The company that builds these models couldn’t make a reliable one and pulled it. That should temper how much weight anyone puts on a percentage from a third-party tool.

So a “likely AI” flag is a statistical hunch about your writing’s texture. The Stanford research shows that hunch is biased — not maliciously, but structurally — against people writing in a second language. If your honest work is flagged, you’re allowed to say so, bring your drafts, and ask a human to actually read the writing. You can also browse more guides on AI detection and writing to get specific on how to make that case.

Frequently asked questions

Do AI detectors actually discriminate against ESL students?

Not by intent, but the effect is real and measurable. The 2023 Stanford study found seven common detectors misclassified an average of about 61% of non-native English essays as AI-generated, while flagging almost none of the native US student essays. The tools aren’t programmed to target anyone, but because they rely on statistical predictability, they penalize the steady, lower-perplexity style common in second-language writing.

Why does my human-written essay get flagged as AI?

Most detectors estimate how predictable your word choices are (perplexity). Clear, simple sentences with a smaller working vocabulary look statistically predictable, which is the same signal a model produces. The detector can’t tell careful plain writing apart from machine writing, so it guesses wrong.

Is it “cheating” to revise flagged writing so it reads more naturally?

No. Editing your own genuine work for clearer, more varied phrasing is just editing. It’s only an ethics problem if you’re passing off generated text as your own. Improving the rhythm and word choice of writing you actually authored is normal revision.

What should I do if a teacher accuses me based on a detector score?

Stay calm and ask what the score means. Detector outputs are probabilities, not evidence. Bring your drafts, notes, and version history, and ask for a human review of your process. Citing the Stanford research on ESL false positives is fair and factual — many instructors don’t know detectors carry this known bias.

The bottom line

If you write in English as a second language and a detector calls your honest work “AI,” the data says the tool is likely wrong, not you. The bias is built into how perplexity-based detection works, and the fix isn’t to write like someone else — it’s to write clearly, vary your rhythm, and keep proof of your process. Want to see how your own draft reads through that statistical lens before someone else does? Run it through PaperBleach and find out where you stand.