Why Graduate-Level ESL Writing Gets Flagged More Than Undergraduate Work

A student writes a sharp, careful argument essay in their first year and the AI detector waves it through. Three years later the same student, now in a master’s program, submits a literature review and gets pulled into the academic integrity office. Their English didn’t get worse. Their writing got more advanced. And that, strangely, is the problem.

If you write English as a second language and you’ve moved up into graduate work, you may have noticed the flags getting more frequent, not less. Here’s why that happens, and what you can actually do about it.

Key takeaways

• Graduate ESL writing carries two penalties at once: the predictable, high-frequency vocabulary common in second-language writing, *and* the formal, hedged, uniform register that graduate programs explicitly reward.
• Most detectors measure perplexity and burstiness, not authorship. Polished academic prose scores low on both, so it can read as machine-made even when no machine touched it.
• A loose undergrad reflection reads as “human.” An impersonal methods section reads as “AI.” The grading rubric and the detector want opposite things.
• A 2023 Stanford study found seven popular detectors misflagged non-native English essays as AI 61% of the time on average. The register problem stacks on top of that.
• The fix isn’t to dumb down your work. Keep your version history, write in your real voice inside the formal frame, and treat any score as a signal, not proof.

Two penalties, stacked

There’s a known, measured bias in AI detectors against non-native English writers. A 2023 team led by Weixin Liang, publishing in the journal *Patterns*, ran non-native TOEFL essays and native-speaker essays through seven widely used GPT detectors. The detectors were near-perfect on the native writing and wrong more than half the time on the non-native writing — 61.22% misflagged as AI on average. The likely culprit, the authors said, was low perplexity: second-language writing tends to use fewer rare words and less lexical variety, and that’s the same statistical fingerprint AI text leaves.

So that’s penalty one, and it follows ESL writers from their first essay to their dissertation.

Penalty two is what graduate school does to your prose on purpose. Undergraduate writing is often personal and argumentative — a reflection, an opinion piece, a close reading where your own take is the point. Graduate writing pushes the opposite direction. A literature review is supposed to sound measured and impersonal. A methods section is supposed to be repeatable and flat. You’re trained to hedge (“these results suggest,” “it may be the case that”), to stay consistent in terminology, to keep your sentences orderly. That’s good academic discipline. It’s also a recipe for low perplexity and low burstiness.

For ESL writers, the two penalties compound. You’re already reaching for safe, dependable words because you’re writing in a learned language. Then your program asks you to write in the one register that flattens variation even further. The result is prose that detectors read as suspiciously machine-like — not because you cheated, but because you did exactly what was asked.

What the tools are actually looking at

It helps to know the mechanism, because once you see it the pattern stops feeling random.

Most statistical detectors revolve around two numbers. Perplexity measures how surprised a language model would be by your word choices — predictable words mean low perplexity. Burstiness measures variation in sentence length and rhythm — humans tend to mix a short punchy line with a long winding one, while models often hold a steadier pace. GPTZero, for example, describes its approach around exactly these ideas. We’ve written a fuller plain-English walkthrough of perplexity and burstiness if you want the mechanics.

Now line that up against a graduate methods section. Standardized terminology, repeated sentence frames (“Participants were recruited via… Data were collected using… Responses were coded according to…”), no slang, no idioms, no rhetorical swerves. Low perplexity. Low burstiness. To a detector, that paragraph looks a lot like generated text. The detector isn’t malfunctioning. It’s measuring the wrong thing — surface predictability — and calling it authorship.

A quick scenario: meet Daniyar

Let me make it concrete. The numbers below are illustrative, not from the study, but the pattern matches the research.

Daniyar is finishing a master’s in public health. Kazakh and Russian are his first languages; English is his third. As an undergrad he wrote argumentative essays, and his professors liked his clear, direct style. His essays scored low on the detector his school used — 5%, 8%, nothing alarming.

Now he’s writing a systematic review. The genre demands a flat, careful voice: he summarizes twelve studies, hedges every claim, keeps his terminology identical across pages so a reader never gets confused about whether “incidence” and “prevalence” mean the same thing. He writes every word himself, over two weeks, with a stack of annotated PDFs beside him.

The detector returns 84% likely AI.

Nothing in his review is wrong. The discipline he learned — consistency, restraint, precision — is the discipline a model imitates. A classmate who writes looser, with the occasional aside and an uneven rhythm, scores 11%. She didn’t write a more honest review. She wrote a *burstier* one. The number can’t see Daniyar’s two weeks, his PDFs, or his honesty. It sees predictability, and at the graduate level he was rewarded for producing more of it.

Why this feels worse than undergrad — and sometimes is

A few things make the graduate experience sharper than the undergraduate one:

• The genres skew formal. Reviews, proposals, and methods sections dominate, and those are the lowest-perplexity, lowest-burstiness writing you’ll ever do. First-person reflection, which reads as human, mostly disappears.
• The stakes are higher. A flagged undergrad essay might mean a redo. A flagged thesis chapter can stall a degree, a visa, or a funding line. International graduate students often carry the most risk for the writing most likely to be flagged.
• Discipline jargon cuts both ways. Heavy field-specific vocabulary can *raise* perplexity in a narrow technical passage, but the connective tissue around it — the framing, the transitions, the hedges — stays flat and predictable. Detectors often weight that connective prose heavily.

None of this means graduate ESL writers are doomed. It means the deck is stacked in a specific, understandable way, and understanding it is the first defense.

What you can actually do

You can’t rewrite the detectors. You have more control over your own position than you might think.

Keep a process record. This is the strongest protection by a wide margin. Write in Google Docs or Microsoft Word so version history saves automatically. A chapter that grew over days — with edits, abandoned paragraphs, reorganized sections — is very hard to fake and very persuasive to a committee. Keep your reading notes and outline drafts too.

Write in your real voice inside the formal frame. After reading about this bias there’s a temptation to inflate your vocabulary to dodge the perplexity trap. The Liang team showed it works on the tools — prompting a model to rewrite flagged essays with fancier wording dropped the false-AI rate from about 61% to about 12%. But that mostly proves the tools react to surface word choice, not authorship. Faking a voice you don’t have can muddy your meaning and look strange to an advisor who knows your writing. Vary your sentence lengths the way you naturally would. Let one short sentence sit next to a long one.

Anchor every section in specifics only you have. Generic, averaged-out prose is what models produce. The exact instrument you used and why you rejected the alternative. The odd result that didn’t fit your hypothesis. A limitation you noticed at 1 a.m. Detail no one could invent for you is the opposite of low-perplexity filler.

Use a checker to understand your own draft, not to game anyone. Running your own work through a detector before you submit is about literacy. If a section you wrote yourself lights up, that’s useful information about which parts of your natural register trip the wire. You can then decide whether to adjust or simply keep your receipts. If you do this regularly across a thesis, it’s worth knowing what the paid tiers cover — see the plans before you commit.

It’s worth remembering how uncertain these tools are even by their makers’ admission. OpenAI launched its own AI text classifier in early 2023 and shut it down that July for low accuracy — its own numbers had it catching only about 26% of AI text while flagging human writing 9% of the time. A detector score is a probability a vendor publishes, not a verdict. Treat it accordingly.

Frequently asked questions

Why would my master’s thesis get flagged when my undergrad essays never did?

Because the writing you do at the graduate level is supposed to sound the way detectors expect AI to sound. Literature reviews, methods sections, and discussion chapters reward an impersonal, hedged, evenly structured register. That produces low perplexity (predictable word choices) and low burstiness (little variation in sentence rhythm), the exact statistical signals most detectors key on. Undergraduate work, especially reflections and argument essays, tends to be more personal and uneven, so it reads as more human. If you also write English as a second language, the careful, high-frequency vocabulary common in second-language writing compounds the effect.

Is there research showing detectors are biased against non-native English writers?

Yes. The 2023 Liang study, published in *Patterns*, ran non-native TOEFL essays and native-speaker essays through seven widely used GPT detectors. The detectors wrongly flagged the non-native essays as AI 61.22% of the time on average while staying near-perfect on native-speaker writing. The authors point to low perplexity, driven by less lexical variety, as the likely cause. Graduate academic register tends to lower perplexity even further, which is why the problem can feel worse at that level.

Should I add complexity or slang to avoid being flagged?

Be careful here. The Liang team found that prompting a model to rewrite flagged essays with fancier vocabulary dropped the false-AI rate from about 61% to about 12%, so dressing the language up does fool the tools. But that mostly proves the tools react to surface word choice, not authorship. Faking a voice you don’t have can bury your meaning and look odd to an advisor who knows how you write. The honest move is to vary your sentence lengths naturally, keep the specific details only you would know, and stay in your real voice.

What evidence protects me if a detector flags my graduate work?

Process evidence is the strongest defense. Write in Google Docs or Microsoft Word so version history is captured automatically; a document that grew over days, with edits, dead ends, and reorganized sections, is hard to fake. Keep your notes, reading annotations, and outline drafts. If you’re flagged, ask which tool was used and what score it returned, then point out that detectors output a probability rather than proof and that published research documents bias against non-native writers. Lead with the record, not an apology.

Where this leaves you

The graduate-versus-undergraduate gap isn’t a mystery once you see the mechanism. Advanced academic writing flattens the exact signals detectors use to guess at authorship, and second-language writing flattens them further. The writers least likely to be cheating end up writing the prose most likely to be flagged.

So protect your position. Keep your drafts and version history. Write like you, within the formal frame your field demands, and load every section with detail only you could supply. And if you’re curious about the wider mechanics, you can always browse the rest of our writing on detection.

Want to see how your draft reads to a detector before a committee does? Check and humanize your draft with PaperBleach.