The Limited Vocabulary Trap: Why Smaller Word Range Makes ESL Writing Read as AI

You can write flawless grammar, hit every point in the prompt, and still get a paper flagged as AI. If English is your second language, the reason often hides somewhere you’d never look: not in your mistakes, but in the size of the word range you’re working from. A smaller vocabulary makes you lean on the same safe words, and that habit produces exactly the pattern detectors are tuned to catch.

This isn’t about writing badly. It’s a quiet collision between how second-language writers build sentences and how detection tools read them.

Key takeaways

• A smaller active vocabulary forces you to reuse a handful of safe, common words, and that low variety is one of the surface signals AI detectors react to most.
• Detectors don’t measure how good your vocabulary is. They measure how *predictable* each word is, and the most common word is also the most predictable one.
• ESL writers get caught twice: limited range pushes word choice toward the predictable, and translation habits pull whole phrases toward standard, textbook English.
• The fix isn’t bigger words. It’s more specific words, controlled repetition, and concrete detail only you could write.
• A high score is a probability, not proof. OpenAI retired its own classifier in July 2023 for low accuracy, and a 2023 Stanford study found detectors biased against non-native English writers.

What detectors actually count when they look at your words

Most AI detectors don’t understand your essay or read it the way a teacher does. What they do is estimate how *surprising* each word is, given the words around it. That surprise level has a name, perplexity, and it sits at the core of how detection works. We break the math down in our piece on the statistics behind AI detectors, but the short version matters here.

Low surprise looks machine-made. AI models are trained to pick the safest, most likely next word, so their text is smooth and predictable. A human with a deep, comfortable vocabulary tends to wander. They reach for the odd verb, the specific noun, the word that’s a little off the beaten path. That wandering reads as high perplexity, and detectors read it as human.

Here’s the catch. If your active vocabulary is smaller, you can’t wander as much. You stick to the words you trust. And the words you trust are usually the common ones, the same ones the AI would have picked.

Range versus repetition

Two things are getting tangled, so let’s separate them.

Your vocabulary *range* is how many words you can reach for at all. Your vocabulary *repetition* is how often you actually reuse the same ones in a single piece. Linguists have a rough measure for the second thing, the type-token ratio: count the unique words, divide by the total words. Lots of repetition means a low ratio.

Detectors react to what’s on the page. They see repetition and low variety. They don’t see how big your mental dictionary is. That’s actually good news, because repetition is something you can fix in editing even if your range is still growing.

Why ESL writers fall into this trap harder

There are two habits stacking on top of each other, and most non-native writers don’t notice either one.

The first is the safe-word reflex. When you’re not sure a word is right, you don’t risk it. You use *important* instead of *pivotal*, *show* instead of *underscore*, *good* instead of *deft* or *shrewd*. Every one of those safe choices is a high-probability word. String enough of them together and your text gets smooth in exactly the way a model’s output is smooth.

The second is translation drift. If you draft a sentence in your first language and carry it into English, the English version tends to snap to the most standard, textbook phrasing. There’s nothing wrong with standard phrasing. But “standard” and “most predictable” are nearly the same thing to a detector. Your sentences land on the expected word over and over, and the variety that signals a human hand never shows up.

This is measured, not theoretical. A 2023 study led by Weixin Liang at Stanford ran non-native TOEFL essays through seven popular detectors and found they were flagged as AI far more often than native-speaker essays, even though every essay was human-written. The researchers tied it to low perplexity, the same safe, predictable word choice we’re describing. It’s a structural blind spot, not a judgment about whether your writing is any good.

A mini-scenario: two ways to write the same sentence

Say you’re writing about a marketing campaign for class. Here’s a version built from safe, repeated words:

> The campaign was very successful. It was successful because it used social media in a good way. Social media is a very important tool for marketing today.

Look at the repetition: *successful* twice, *social media* twice, *very* twice, and every word is the most expected one for its slot. A detector sees almost no surprise here: low variety, low perplexity, high AI score, even though a stressed human at 1 a.m. wrote it.

Now the same idea, with precise words and one concrete detail:

> The campaign worked. Specifically, a 12-second TikTok clip of our product failing on purpose pulled in 40,000 views overnight, mostly because people love watching something break before it works.

Nothing fancy, no thesaurus words. But the nouns got specific, one verb did real work, and there’s a detail no model would have guessed. The variety went up because the *precision* went up, not because the words got harder.

How to fix it without faking a voice that isn’t yours

The wrong instinct is to grab a thesaurus and swap in big words. Don’t. Words you don’t fully control read as forced to a human grader, and an essay that sounds like a stranger wrote it is a worse problem than any detector score. The Stanford team did find that enriching word choices lowered the false-AI rate, but that mostly proves the tools react to surface word choice, not to who wrote the piece. Gaming the surface is fragile. Here’s what holds up.

Reach for the specific noun, not the bigger one. “The thing improved” is weak because *thing* and *improved* are maximally generic. “Load times dropped from 9 seconds to 3” is strong because it’s exact. Specificity raises variety as a natural side effect.

Hunt your repeats in the second draft. Read through once looking only for words you used twice in a paragraph. Replace the second one with something precise, or restructure the sentence so you don’t need it. This single pass does more for your lexical variety than any other habit.

Add one detail only you could know. A date, a place, a number from your own work, a small thing that went wrong. Models generalize. Humans remember the weird specific. That texture is the hardest signal for a detector to fake away.

Vary length on purpose. A three-word sentence after a long one breaks the steady rhythm detectors also flag. Pair that with word variety and the whole piece reads more human.

If you want to see where your draft stands before you hand it in, you can run your draft through PaperBleach and watch which sentences light up, then fix those instead of guessing.

One thing to keep in perspective

A flag is not a verdict. AI detectors output a probability about word patterns, not proof about who typed them. OpenAI shut down its own AI Text Classifier in July 2023, citing a low rate of accuracy, and scores still drift between tools and sometimes between runs on the same text. So if you write in a second language and get flagged, keep your drafts, notes, and version history. Your process is the strongest evidence you have, and a detector can’t see it. For more on the wider picture, you can browse the rest of our writing on detection.

Frequently asked questions

Does a small vocabulary really make my writing look AI-generated?

It contributes, yes. A limited active vocabulary means you reach for the same handful of safe, high-frequency words again and again, because those are the ones you trust. Detectors estimate how predictable each word is, and the most common word in a slot is also the most predictable one. So writing built from safe, repeated choices produces the low-surprise pattern these tools associate with machine text. It’s not that your vocabulary is too small to be human. It’s that cautious, repetitive word choice overlaps statistically with an AI playing it safe.

Should I just use bigger or fancier words to avoid being flagged?

No, and this is the trap most people fall into. Swapping in thesaurus words you don’t control usually makes your writing worse and reads as forced to a human grader, which is a bigger problem than a score. The honest, durable fix is precision and specificity, not vocabulary cosplay. Use the exact word for the thing, add a concrete detail, and let your real range show.

Why are non-native English writers hit by this more than native speakers?

Two habits stack. A developing vocabulary leans on common, safe words, which are the predictable ones. And composing by translating from your first language snaps whole phrases to standard, textbook English, clustering your word choices around the most expected option. The 2023 Stanford study found detectors misclassified non-native essays as AI far more often than native-speaker essays. The tools weren’t built to target ESL writers; the way they measure text just penalizes normal second-language patterns.

If my essay gets flagged, does that prove I used AI?

No. A detector outputs a probability about word patterns, not evidence about who typed them. OpenAI retired its own classifier in July 2023 for low accuracy, and scores drift between tools and even between runs. A flag is a reason to look closer, never a verdict. Keep your drafts and version history so you can show your process.

What’s the difference between a limited vocabulary and a repetitive one?

They overlap but aren’t identical. A limited vocabulary is about range, how many words you can reach for. A repetitive one is about how often you reuse the same words in a given piece. You can have wide range and still write repetitively under time pressure. Detectors react to repetition on the page, not to how big your mental dictionary is, which is why editing for repeats works.

Closing

A smaller word range isn’t a flaw or a sign you cheated. It’s just a habit that shares a fingerprint with cautious AI, and you can edit around it without pretending to be someone you’re not. Write the specific word, hunt your repeats, and drop in the detail only you would know. To spot the risky sentences before a teacher’s tool does, see the plans and fix what actually lights up.