Idioms, Collocations, and Why Avoiding Them Makes ESL Writing Look Machine-Made

You learned, somewhere along the way, that the safe move in English is to keep it simple. Don’t risk an idiom you might get wrong. Reach for the plain word, the textbook phrasing, the construction you’re sure about. It feels careful and correct—and that’s exactly the problem. That same caution can make your writing read as machine-made to an AI detector.

This isn’t a knock on ESL writers. It’s a quirk of how detectors work and how careful writing tends to flatten. Once you see it, you can fix it without pretending to be someone you’re not.

Key takeaways

• Collocations are the word pairings native speakers expect—”make a decision,” not “do a decision.” Avoiding them flattens your prose into safe, generic phrasing.
• AI detectors react to predictability. Both over-cautious ESL writing and machine output can look statistically smooth and low-surprise.
• Skipping idioms and natural pairings removes the small irregularities that make human writing feel human.
• The fix isn’t cramming more idioms. It’s letting natural collocations and a few rough edges back into your drafts.
• Detector scores are probabilities, not proof. A flag means “this looks predictable,” not “a machine wrote this.”

What collocations actually are

A collocation is a set of words that English speakers habitually put together. “Heavy rain” sounds right. “Strong rain” doesn’t, even though strong and heavy are near-synonyms. You “make a decision,” you don’t “do a decision.” Coffee is “strong,” not “powerful.” None of this is grammar in the strict sense—the alternatives parse fine. It’s convention, built up over millions of real conversations and pages of text.

Native speakers absorb these pairings without noticing. ESL writers often learn the words but not the company they keep. So an advanced learner with great grammar can still produce sentences that feel subtly off, because the individual words are right but the combinations aren’t the ones a reader expects.

Idioms are the louder cousin of collocations—”bite the bullet,” “cut corners,” “on the fence.” They carry meaning you can’t work out from the individual words. Both idioms and collocations do the same quiet job: they make writing sound lived-in.

Why playing it safe backfires

Here’s the trap. When you’re not sure which pairing is correct, the rational move is to avoid the risk. Instead of “the policy cut corners on safety,” you write “the policy did not include enough safety measures.” Instead of “demand skyrocketed,” you write “demand increased significantly.” Each swap is defensible. Stack a few hundred of them across an essay and the text turns smooth, generic, and predictable.

That predictability is the issue. AI detectors don’t read for meaning. They estimate how surprising your word choices are to a language model. If your next word is almost always the most expected one, the text scores as low-surprise—and low-surprise is the signature detectors associate with machine generation. We walk through the math in perplexity and burstiness, but the short version: the careful, high-probability vocabulary that feels safe to you looks like a machine to the detector.

So you end up in a frustrating spot. You wrote every word yourself. You avoided idioms precisely because you wanted to be correct. And the detector flags you anyway, because correctness without texture reads as predictability.

A mini-scenario

Two students summarize the same reading.

Priya, a confident native speaker, writes: “The author bends over backwards to stay neutral, but you can feel where she lands.” It’s loose, a little informal, full of natural pairings.

Mateo, a careful ESL writer, writes: “The author makes a significant effort to remain neutral, but it is possible to identify her position.” Cleaner grammar, honestly. Every word is correct. But it’s also the most expected version of nearly every phrase—”makes a significant effort,” “it is possible to,” “identify her position.” A detector sees a low-surprise sentence and nudges its score up.

Mateo did nothing wrong. He played defense, and defense reads as flat.

This isn’t your fault—and the bias is real

It’s worth saying plainly: the deck is somewhat stacked. In 2023, a study led by Stanford’s James Zou and published in the journal *Patterns* tested seven popular detectors on essays by non-native English speakers. The detectors flagged those essays as AI-generated far more often than they flagged writing by native speakers. The reason lines up with everything above—non-native writing tended to use simpler, more predictable vocabulary, and the detectors read that predictability as a machine fingerprint.

That same year, OpenAI retired its own AI text classifier, citing a low rate of accuracy. The tools keep improving, but they remain probability estimators, not lie detectors. A score is a guess about how predictable your text looks. It is not evidence that a machine wrote it.

Keep that framing. The goal isn’t to trick anything. It’s to write English that sounds like you actually wrote it—which, conveniently, also happens to be less predictable.

How to let the texture back in

The fix is not memorizing a list of 500 idioms and sprinkling them on top. Forced idioms read worse than plain language; nothing flags faster than “it’s raining cats and dogs” wedged into an economics paper. The real work is more natural than that.

Read and listen to real English in your field

Collocations live in context. Read articles, watch talks, and skim forum threads in the area you write about. Notice which words travel together. Economists say demand “softens” and prices “edge up.” Programmers “ship” features and “patch” bugs. You can’t reason your way to these pairings—you collect them by exposure, the same way native speakers did, just on purpose.

Trust the pairings you already know

A lot of ESL writers know more natural English than they use, because they’re busy censoring themselves. If “demand skyrocketed” comes to mind first, that instinct is probably right. Stop overriding it with the cautious version. Let the first natural phrasing stand more often.

Read your draft out loud

This is the cheapest tell-detector you have. Flat, over-careful writing sounds monotone when spoken. If every sentence lands at the same length and the same register, you’ll hear it. Break the rhythm. Let one sentence run long and the next be three words.

Use specifics instead of safe abstractions

“Increased significantly” is safe and empty. “Doubled in six months” is concrete and unexpected. Specifics naturally lower predictability because the exact number or example isn’t something a model would guess. Reaching for the concrete detail does double work: stronger writing, and writing that looks less robotic.

Check, then adjust by hand

Before you submit, it helps to see which sentences a detector reads as predictable, then treat those flagged spots as a prompt to add texture—a natural collocation, a concrete number, a varied sentence length. Don’t treat a flag as a verdict on your character. The point is to learn what flat looks like so you can hear it yourself next time.

A worked example

Take Mateo’s sentence again: “The author makes a significant effort to remain neutral, but it is possible to identify her position.”

Now revise it with natural pairings and a specific touch: “The author works hard to stay neutral, but her stance leaks through—especially in the third section, where she calls the policy ‘reckless.'”

Notice what changed. “Works hard to stay neutral” is a common collocation, not a textbook construction. “Leaks through” is a small, vivid pairing. And the specific detail—the third section, the word “reckless”—gives the sentence something a model wouldn’t predict. It’s still clean, still correct, still Mateo’s own thought. It just sounds like a person made it.

Frequently asked questions

What is a collocation, exactly? A collocation is a group of words English speakers habitually use together—”heavy rain,” “make a decision,” “strong coffee.” The grammar of an alternative might be fine (“powerful rain”), but it sounds off because the conventional pairing is different. Speakers learn these by exposure, not rules, which is why they trip up even advanced ESL writers.

Why would avoiding idioms make my writing look AI-generated? When you strip out idioms and natural pairings, you fall back on the most generic, highest-probability words. That produces smooth, predictable text with little variation—exactly the statistical pattern many detectors associate with machine output. The carefulness reads as flatness.

Do AI detectors actually penalize non-native English? Not on purpose, but research has shown bias in practice. The 2023 study led by Stanford’s James Zou, published in *Patterns*, found that seven popular detectors misclassified non-native TOEFL essays as AI far more often than native writing, likely because that writing used simpler, more predictable vocabulary. Detectors measure predictability, and over-careful ESL prose can be very predictable.

Should I just memorize a list of idioms? Memorizing helps a little, but forced idioms often read worse than plain language. Better: read and listen to a lot of real English in your field, notice which words travel together, and let those pairings into your drafts. Natural collocations first; idioms are optional seasoning.

If my essay gets flagged, does that mean I’ll be accused of cheating? Not by itself. A detector score is a probability estimate, not evidence. A high score means your text looks predictable, which can happen for honest reasons including ESL phrasing. Keep your drafts, notes, and sources so you can show your process if anyone asks.

The bottom line

Careful writing and machine writing can land in the same place: smooth, predictable, and oddly impersonal. For ESL writers, the path out isn’t more caution—it’s a little more of yourself. The natural pairing you almost used. The concrete number. The sentence that runs long because the thought was long. That texture is what reads as human, to people and to detectors alike.

Want to see where your own draft sounds flat before you turn it in? Run it through PaperBleach and use the flagged sentences as a map for where to add a little life. For more on this, browse our other articles on AI detection and ESL writing.