PaperBleach for ESL Writers: How Humanizing Differs When English Is Your Second Language

If you learned English as a second language, you’ve probably had this stomach-drop moment: you write an essay yourself, word by word, and a detector tells you it’s “98% AI.” You didn’t cheat. The tool just doesn’t know the difference between careful AI text and careful non-native English. That gap is real, it’s measurable, and it changes how humanizing should work for you.

Key takeaways

• AI detectors lean on perplexity and burstiness, two patterns that often look the same in polished AI text and in careful non-native English.
• A 2023 Stanford study (Liang et al., in the journal Patterns) found GPT detectors flagged more than half of TOEFL essays by non-native speakers as AI, while almost never flagging native-speaker essays.
• For ESL writers, humanizing isn’t about hiding a machine. It’s about adding the natural unevenness detectors read as human, without erasing your meaning.
• A good humanizer should preserve your argument and word choices, not flatten everything into generic native-sounding filler.
• The goal is honest writing that reads naturally, not a “guaranteed” score. Detectors give probabilities, not proof.

Why ESL writing trips the same wires as AI

Most detectors don’t read your essay for meaning. They measure two things. First, perplexity: how surprising your word choices are to a language model. Second, burstiness: how much your sentence length and rhythm vary. AI text tends to be low-perplexity (it picks the most likely next word) and low-burstiness (it keeps an even, steady pace).

Here’s the uncomfortable overlap. When you write in a second language, you naturally play it safe. You reach for the common, “correct” word instead of the unusual one. You build sentences in a steady, structurally similar way because that’s what you’ve practiced and trust. The result reads as low-perplexity and low-burstiness too. To the detector’s math, your honest effort and a machine’s output look nearly identical.

If you want the full mechanics of how this scoring works, we wrote a plain-English breakdown of perplexity and burstiness — the two statistics behind AI detectors. The short version: the very habits that make you a disciplined non-native writer are the habits that push your score up.

This isn’t a hunch — it was measured

In 2023, a Stanford team led by Weixin Liang tested several popular GPT detectors on essays written by non-native English speakers (TOEFL exam essays) and on essays written by native-speaking US students. The detectors flagged more than half of the non-native essays as AI-generated, while their accuracy on the native-speaker essays was near-perfect. Same task, wildly different error rates, split along language background. The work was published in the journal Patterns.

The reason mapped to what we described: the non-native essays scored lower on lexical “richness” and perplexity, and the detectors treated that as an AI signal. None of those students used AI. The tool was biased by design, not by accident.

It’s worth remembering the broader context too. OpenAI quietly retired its own AI Text Classifier in July 2023 because its accuracy was too low to be useful. If the company building the models can’t reliably detect them, a single detector score against your essay deserves real skepticism.

So humanizing means something different for you

For a native English speaker, the usual problem with AI text is the opposite of yours. Their drafts are often too polished, too even, too relentlessly parallel — they need to rough up the rhythm. As an ESL writer revising your own work, you’re frequently starting from a place that’s already even and careful. The fix isn’t to “sound more like a robot’s opposite.” It’s to introduce the natural variation that your writing habits tend to smooth out, while protecting the meaning you fought to get right.

That last part matters more for you than for almost anyone. The biggest danger of a careless humanizer is that it treats your precise vocabulary as a problem to “fix.” You chose “mitigate” on purpose; a crude tool swaps it for “lessen” and quietly changes your nuance. Multiply that across a paragraph and your argument blurs.

A quick scenario

Imagine Mariana, a Brazilian grad student writing a public-health discussion post. Her draft:

> “Vaccination programs reduce disease transmission. Public trust is important for participation. Governments should communicate clearly. This approach improves outcomes.”

Every sentence is true, grammatical, and roughly the same length and shape. A detector reads that flat, even rhythm and low surprise and returns a high AI probability — even though Mariana typed every word herself.

Good humanizing for Mariana doesn’t rewrite her ideas. It varies the structure and reintroduces the texture a steady second-language rhythm tends to flatten:

> “Vaccination programs cut disease transmission — that part’s settled. What’s harder is participation, and participation depends on trust. When governments communicate clearly and early, people show up. When they don’t, no amount of program design makes up for it.”

Same argument. Same vocabulary level. But now there’s a long sentence next to a short one, a dash, a fragment of contrast. The meaning is hers; the rhythm just stopped looking machine-flat.

How PaperBleach handles this without erasing you

The thing to avoid is a tool that rewrites your whole document into one generic “native” voice. That solves the detector and destroys the writer. PaperBleach works differently, and it matters specifically for non-native writers.

It scores your text with a sentence-level heatmap, so you can see *which* sentences read as AI-like and which already read as human. Then it rewrites only the flagged sentences and leaves the rest alone. For an ESL writer, that means the careful, correct passages you’re proud of stay yours — only the spots a detector actually objects to get touched.

A practical workflow:

1. Write the whole thing yourself first. The ideas, the evidence, the structure — all you. Don’t outsource thinking.
2. Run it and read the heatmap. Notice the pattern. Often it’s not your “bad English” getting flagged; it’s your *most even, most correct* paragraphs.
3. Let it revise the hot sentences, then reread every change. This is non-negotiable for ESL writers. If a rewrite swapped a word that carried specific meaning, put your word back. The tool serves your voice, not the other way around.
4. Check the score again, and stop when it’s reasonable. Don’t chase a perfect zero. Natural human writing isn’t a zero either.

If you write regularly and want to know what a plan covers, you can see how the plans compare. And when you’re done here, there’s a whole library of more articles on AI detection and writing that go deeper on specific cases.

Keep your editing honest

Humanizing your own writing so it reads naturally and survives a biased detector is fair editing. Generating an essay you never thought through and passing it off as your analysis is not — and no tool changes that line. As an international student or non-native professional, you may face stricter scrutiny, so protect yourself: keep your drafts, your notes, your outline, your version history. If a detector ever flags your genuine work, that paper trail is your strongest defense, because the score itself is only a probability.

Frequently asked questions

Why do AI detectors flag non-native English writers so often? Detectors estimate how predictable your text is (perplexity) and how much your sentence length varies (burstiness). Non-native writers tend to use safer, common words and steadier rhythms, which lowers both — the same statistical fingerprint detectors associate with AI. The 2023 Stanford study by Liang and colleagues, published in Patterns, documented this bias directly.

Is using an AI humanizer as an ESL writer cheating? Not on its own. If the ideas and argument are genuinely yours and you’re revising your own writing to read more naturally or to avoid a false positive, that’s editing. The line you shouldn’t cross is submitting work you didn’t actually think through. Always follow your school’s or client’s disclosure rules.

Will a humanizer make my writing sound less like me? It can, if it’s a blunt synonym-swapper — that’s the main risk for ESL writers. PaperBleach rewrites only the sentences a detector flags and leaves the rest alone, which keeps more of your voice. Still, reread the output and restore any phrasing that mattered.

What’s the difference between humanizing for ESL writers and for native speakers? Native speakers usually need to break up over-polished, repetitive AI rhythm. ESL writers more often need to add natural variation and flow without losing meaning. Different starting points, different editing focus.

Can my real, self-written essay get flagged as AI? Yes, and it happens to non-native writers more than anyone. A high score isn’t proof — it’s a probability. If you wrote it yourself, keep your drafts and version history as evidence.

The takeaway

You’re not the problem. The detector’s math is. As a second-language writer, your careful, correct habits happen to look like the thing these tools are built to catch — which is why humanizing, done right, means adding natural texture without losing the meaning you worked to express. Want to see exactly which sentences a detector objects to in your own writing? Try PaperBleach on your next draft.