ESL Students and AI Detectors: Why You Get Flagged and What to Do

You did the work. You wrote the essay yourself, in English you’ve spent years learning, and a detector still told your professor it was probably made by a machine. If that’s happened to you, you’re not imagining things, and you’re not alone — this is one of the most repeatable failures in the whole AI-detection world.

Here’s the part that should make you feel better: the flag usually says more about how these tools work than about your writing. Let’s break down why it happens, then walk through exactly what to do — both before you submit and after you’ve been accused.

Key takeaways

• AI detectors don’t read for meaning. They read for statistical patterns, and clean, careful second-language writing happens to match the pattern they link to AI.
• A 2023 Stanford study found that seven popular detectors wrongly flagged non-native English essays as AI more than 60% of the time — over half — while barely missing on native-speaker essays.
• Prevention beats damage control. Keep a visible writing trail, and let your real voice and specific details show up on the page.
• A detector returns a probability, not proof. OpenAI even pulled its own classifier in July 2023 for low accuracy.
• If you’re flagged, lead with evidence and questions, not apologies.

What the detector is actually measuring

Start with the most important misunderstanding. An AI detector is not a lie detector, and it can’t see whether you opened a chatbot. It looks at your finished text and asks a narrow statistical question: *how predictable are these word choices?*

AI models build sentences by repeatedly picking safe, high-probability words. That makes their output statistically “smooth.” Detectors learn to treat that smoothness as a sign of machine writing. If you want the full mechanics, we’ve covered the statistics behind how detectors score text in plain language elsewhere.

Now think about how you write in a language you learned later in life. You reach for words you’re confident are correct. You skip slang you’re not 100% sure about. You keep sentences clean because being understood matters more than sounding clever. Every one of those is a good instinct — and every one of them pushes your writing toward the same smooth, predictable surface the detector reads as “AI.”

So the flag isn’t a judgment that your writing is bad. In a strange way, it’s often a side effect of writing *carefully*.

The study that put a number on it

This isn’t a hunch. In 2023, a Stanford team led by Weixin Liang published a paper in the journal *Patterns* called “GPT Detectors Are Biased against Non-Native English Writers.” They ran two sets of essays through seven widely used detectors: TOEFL essays written by non-native English speakers, and essays written by US students who were native speakers.

The detectors were nearly perfect on the native-speaker essays — they correctly cleared more than 90% of them. On the non-native essays, they were wrong more than half the time. More than 60% of those human-written TOEFL essays got tagged as AI, and one detector flagged nearly all of them (Liang et al., 2023). Same task, same tools. The main thing that changed was whether English was the writer’s first language, and the result flipped from reliable to coin-flip.

If a detector has misfired on you, you’re inside that statistic, not outside it.

A quick scenario: Khanh’s reflection paper

Let’s make it concrete. The numbers below are illustrative — I made them up to show the pattern — but the pattern itself matches the research.

Khanh is a first-year business student. Vietnamese is his first language. He writes a three-page reflection on a group project over two evenings, entirely himself. His prose is direct because that’s how he thinks in English: “Our team missed the first deadline. I asked everyone to send updates each day. After that, we finished on time.”

Short sentences. Common words. Clear order. No idioms, no flourishes.

His instructor runs it through a detector. It comes back as very likely AI.

Nothing in Khanh’s paper is wrong. The clarity that makes it easy to read is exactly what the tool reads as robotic. Meanwhile a classmate who wrote a messier draft — full of casual asides, tangents, and odd phrasing — barely registers. He didn’t write a better paper. He wrote a less predictable one, and the detector rewarded the unpredictability, not the quality.

That’s the trap in one picture. Which is why your goal isn’t to “beat” the tool. It’s to protect your work and write in a way that’s genuinely, recognizably yours.

What to do *before* you submit

Most of your leverage is here, before anyone runs a scan.

Keep a writing trail you can show

This is the single most powerful protection, and it costs you nothing. Write in Google Docs or Microsoft Word, which both keep automatic version history. A document that grew over hours and days — with edits, deletions, and dead ends — is very hard to argue with. No AI dump looks like that. If you draft in another app, periodically paste your progress into a dated file so you can show the timeline later.

Write in your own voice, not a “smarter” one

After learning about this bias, it’s tempting to inflate your vocabulary to dodge the pattern. The Stanford team showed that dressing essays up with fancier words really did lower the false-AI rate — but that mostly proves the tools react to word choice, not authorship. Faking a more elaborate voice can bury your meaning, and a professor who knows how you write may notice the sudden change. Instead, vary your sentence lengths the way you would in conversation. Pair a short sentence with a long one. Let a little of *you* onto the page.

Add specifics only you would know

Concrete, personal detail is the natural enemy of generic, predictable text. The exact thing your professor said in week three. The name of the lab partner who bailed. A small mistake you made and what it taught you. Models produce smooth 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 machine output — and the more obviously human it reads to an actual person.

Know what a checker costs and what it tells you

Before you submit, it helps to know what a checker can and can’t do, and what a checking and humanizing tool actually offers. A score tells you which sentences happen to trip the detector — useful information about your style, not a verdict on your honesty. If a paragraph you clearly wrote yourself gets flagged, now you know which parts of your natural voice trip the wire, so you won’t be blindsided and you can keep your receipts ready.

What to do *after* you’ve been flagged

If the accusation has already landed, slow down. The instinct to apologize is strong, and it can read as a confession. Don’t.

1. Ask for specifics first. Which tool was used? What exact score did it return? Get the claim pinned down before you respond to it.
2. Present your evidence. Open your version history and show the document growing. Bring your notes, outlines, and earlier drafts. Let the timeline do the talking.
3. Explain what the score is. A detector outputs a *probability*, not proof. Mainstream tools say “likely,” not “definitely,” for a reason.
4. Point to the research, calmly. A peer-reviewed Stanford study documents bias against non-native English writers. You’re not making an excuse; you’re citing a known limitation of the method.
5. Note the tools’ own track record. OpenAI launched an AI text classifier in early 2023 and shut it down that July, citing low accuracy. By OpenAI’s own numbers, the classifier caught only about 26% of AI-written text while falsely flagging human writing 9% of the time. The company that builds these models pulled the plug on its own detector.

Stay factual and steady. You’re not asking for mercy. You’re showing that a single number can’t see two evenings of honest work — and that the number reflects how predictable your word choices were, not who wrote them.

Where this leaves you

The bias against ESL writers isn’t a bug that gets patched next semester. It’s built into how these detectors measure text, and it lands hardest on the writers least likely to be cheating. That’s unfair, and pretending otherwise doesn’t help you.

What helps is control over the parts you own: keep your drafts, write like yourself, and stack up specifics no model could fake. When you want to go deeper on detection and writing well in a second language, you can browse the rest of our writing guides any time.

Try it yourself

Don’t take a detector’s word for it. Paste your draft into PaperBleach’s free AI checker and humanizer to see exactly which sentences get flagged — and make your writing read more like the person who actually wrote it: you.