Reverse Engineering a False Positive: Walking Through One ESL Essay the Detector Got Wrong

You wrote every word. You looked up the connectives your textbook recommended, you hedged your claims the way academic English is supposed to, and you kept your paragraphs tidy. Then the detector came back red, and now you’re staring at your own honest work wondering what gave it away.

The frustrating answer is that nothing “gave it away,” because there was nothing to give. A false positive happens when a genuinely human ESL essay carries the same statistical fingerprint a language model leaves, and this walkthrough of an ESL essay false positive AI detection example shows exactly which sentence habits caused the flag and how to fix them. We’ll take one short paragraph, invented but realistic, and reverse engineer the score line by line.

Key takeaways

• Detectors don’t check who wrote a text. They estimate how predictable your words are and how much your sentence length varies, then guess.
• The habits that make ESL academic writing look “correct” (template connectives, even sentences, reflexive hedges) are the same ones that read as predictable, so honest writing gets flagged.
• Four features usually drive the flag together: low perplexity, uniform sentence length, stacked connectives, and formal hedging on autopilot.
• You can revise the shape without touching your argument: vary your sentence lengths, tie connectives to your real point, and swap empty phrases for specific detail.
• Keep drafts and version history so you can prove your process. A detector score is a probability, not evidence.

An ESL essay false positive: an AI detection example, walked through

Here’s the paragraph. To be clear up front, this is an illustrative example I wrote to demonstrate the pattern. It is not a real student’s essay, not a cited case, and not tied to any specific detector’s output. Treat it as a teaching sample, not a quotation:

> *Nowadays, technology plays an important role in modern education. Moreover, many students use the internet to complete their assignments and to communicate with their teachers. In addition, online learning provides flexibility for students who have busy schedules. It can be said that these developments have changed the way education works. In conclusion, technology has a significant impact on the learning process, and it is widely believed that this trend will continue in the future.*

Read that and nothing feels wrong. The grammar is clean. The register is formal. An examiner skimming it wouldn’t blink. And that’s exactly the problem: it’s so smooth and so expected that a detector rates it as high-probability machine text. Let’s see why.

Sentence by sentence

• “Nowadays, technology plays an important role in modern education.” A stock opener. Every word is the one a model would predict next given the last. There’s no specific claim, no number, no named thing. Pure low perplexity.
• “Moreover, many students use the internet…” A template connective bolted onto the front, and a sentence roughly the same length as the first. The rhythm is already flattening.
• “In addition, online learning provides flexibility…” A third generic signpost in a row. Three sentences, three memorized openers, three near-identical lengths.
• “It can be said that these developments have changed…” A hedge with no one doing the hedging. Said by whom? It’s a reflex, not a judgment, and reflexes are predictable.
• “In conclusion, technology has a significant impact… it is widely believed that this trend will continue.” A formulaic close, an empty intensifier (“significant impact”), and another agentless hedge (“it is widely believed”). The paragraph ends exactly where a model would end it.

None of that is cheating. It’s what happens when you learn to write essays from a phrase bank. But when you add it up, the paragraph has almost no surprise in it, and surprise is precisely what detectors measure.

The four features that drove the flag

If you want the full mechanics, our piece on perplexity and burstiness breaks it down. Here’s how the two ideas show up in the paragraph above.

Low perplexity

Perplexity is how surprised the model is by your next word. In our sample, every word is the most expected one: “important role,” “busy schedules,” “significant impact.” A model would predict each of these, so the detector reads the whole thing as low-perplexity, which it associates with generated text. The fix isn’t fancier vocabulary. It’s saying something specific enough that the next word can’t be guessed.

Uniform sentence length

Burstiness is how much your sentence length varies. Count the words: the five sentences run roughly 9, 17, 12, 11, and 22 words, but they *feel* even because each is one steady medium clause after another, no short punches, no genuinely long build. Human writing under real pressure is lumpier than that. Machine writing tends to be smooth. Smooth is what got flagged.

Template connectives

*Nowadays. Moreover. In addition. In conclusion.* Four sentences, four memorized signposts. The detector doesn’t hate the word “moreover.” What it responds to is the predictable pattern of a generic connective opening line after line. It’s the single most visible tell in ESL academic drafts.

Formal hedging on autopilot

“It can be said that” and “it is widely believed that” are hedges, and hedging is good academic practice. But these hedge nothing in particular. They’re placeholders where a real qualification should go. That reflexive quality is predictable, and predictable is the whole problem.

This is also why detectors misfire on non-native writers as a group. A 2023 study by Liang and colleagues, published in *Patterns*, found several detectors flagged writing by non-native English speakers far more often than native speakers, partly because that writing leans on a narrower band of common, predictable words. And OpenAI quietly retired its own AI text classifier in July 2023, citing low accuracy. These tools guess. Sometimes they guess wrong about honest people.

How to revise those sentences

The goal is not to disguise anything. It’s to write more like a person who actually thought about the topic, which happens to lower the predictability too. Same argument, different shape:

> *Online tools have quietly rewired how my classmates study. Most of us now submit assignments through a portal and message teachers at midnight, which sounds convenient until the wifi drops before a deadline. The real gain is flexibility: a student working a shift can watch the lecture at 6 a.m. instead of skipping it. Whether that flexibility actually improves learning, though, is a separate question, and I’m not convinced the evidence is there yet.*

Look at what changed. Sentence lengths swing from short to long. The connectives are gone, replaced by links carried inside the sentences. “It is widely believed” became a real, personal hedge: *I’m not convinced the evidence is there yet.* A concrete image (the wifi dropping, the 6 a.m. lecture) replaced the empty “significant impact.” The claim is the same. The predictability is not.

You don’t need to do all of this to every paragraph. Fixing the two or three most template-heavy sentences usually moves the score more than a full rewrite, because those were the sentences carrying the flag.

How to protect yourself if you’re flagged

Revising is half the job. The other half is being able to stand behind your work if a tool questions it.

• Keep your process. Draft in something with version history, like Google Docs or Word with autosave on. A document that visibly grew over days is far more convincing than any score.
• Save your outline, notes, and sources. These show the thinking behind the essay. A generator can’t produce your messy pre-writing.
• Self-check before you submit, not after. Run your draft through a detector yourself so a red flag never surprises you. You can check a draft in the detector, see which sentences read as predictable, and revise only those instead of gutting the whole essay.
• Talk to your instructor early. If you know your careful, formal style tends to get flagged, say so before it becomes a dispute. Most will accept a draft history without a fight.

To be honest about the limits here: this isn’t about beating any specific system or sneaking past a checker. It’s about writing that’s genuinely more specific and more varied, and keeping the receipts that prove the words are yours. Those two habits solve the real problem, which is a false positive against work you actually did.

Frequently Asked Questions

What is an example of an ESL essay false positive in AI detection? It’s when a detector rates writing you produced yourself as likely AI, purely because of its statistical shape. A common pattern: an ESL essay built from memorized connectives (Nowadays, Moreover, In addition), sentences that all run about the same length, and reflexive hedges like “it can be said that.” None of that is generated, but it’s low-variation and highly predictable, which is exactly the fingerprint a language model leaves. The detector reads the statistics, not who typed the words, so it guesses AI and gets it wrong.

Why would a detector flag an essay I wrote entirely myself? Because detectors don’t check authorship. They estimate two things: perplexity, or how predictable each word is given the ones before it, and burstiness, or how much your sentence length varies. Learned academic English tends to be low on both, so it overlaps with model output. A 2023 study by Liang and colleagues found detectors flagged non-native English writing as AI far more often than native writing, largely for this reason. A high score is a probability, not proof.

Which sentence patterns most often trigger a false positive? Four repeat offenders: stacked template connectives (Moreover, Furthermore, In addition opening consecutive sentences), a run of sentences all the same medium length, reflexive hedges with no real agent (it can be said, it is widely believed), and empty intensifiers (significant, various, numerous) standing in for a concrete detail. Any one is fine. All four together produce the even, predictable rhythm detectors key on.

How do I revise flagged sentences without changing my meaning? Change the shape, not the claim. Break one long sentence into two and fold two short ones into a longer one so your lengths vary. Replace one or two template connectives with a link tied to your actual point. Swap a reflexive hedge for one that names the real uncertainty, and replace at least one empty intensifier with a specific figure, study, or example. Your argument stays identical; only the predictability drops.

How can I prove I wrote an essay if it gets flagged? Keep evidence of your process. Draft in a tool with version history (Google Docs, Word with autosave) so you can show the essay growing over time. Keep your outline, notes, and sources. Save earlier drafts. If a detector flags your work, that trail is far stronger than any score, and most instructors will accept it. A detector’s output is not evidence of cheating on its own.

The score was never the point

A false positive feels like an accusation, but it’s really just a pattern match, and patterns can be changed without changing the truth of what you wrote. Make your sentences vary. Say specific things. Keep the drafts that show your hand at work. Do that and the red flag stops being a threat, because you’ll have both better writing and the proof to back it.

When you want a second opinion before you hand something in, you can check a draft in the detector and fix only the sentences that still read as predictable. If you’re weighing how often you’ll use it, see which plan fits, or browse more guides on writing and detection to keep the habit sharp.