The Hedging Problem: Why ESL Writers Overuse Phrases That Detectors Associate With AI

You were taught to hedge. Don’t say “this proves” — say “this may suggest.” Don’t say “always” — say “in many cases.” That advice is good, and it’s exactly why your essay just came back flagged as likely AI.

Here’s the uncomfortable overlap: the careful, claim-softening language that good academic writing rewards is also the language AI models generate by default. When an ESL writer leans hard on hedges, the phrasing starts to look statistically like a cautious machine. This post explains why that happens and how to keep your caution without reading as a robot.

Key takeaways

• Hedging — softening claims with phrases like “it could be argued” or “to some extent” — is taught as good academic style, but heavy, formulaic hedging overlaps with the cautious, low-risk phrasing AI models produce by default.
• ESL writers tend to hedge more for understandable reasons: politeness norms carried over from a first language, fear of being wrong in a second language, and textbook templates that supply ready-made hedges.
• Detectors don’t penalize hedging itself; they react to the low-surprise, predictable word patterns that stacked hedges create.
• The fix isn’t to delete all caution — it’s to hedge with specifics: name who argues it, point at the actual limit, and commit to a clear stance where you have one.
• A flag is a probability about phrasing, not proof you used AI.

What hedging actually is

Hedging is the language we use to signal that a claim isn’t absolute. It’s not a mistake. A scientist who writes “results suggest a link” instead of “results prove a link” is being honest about uncertainty, and journals expect it. Linguist Ken Hyland spent a whole book on how central hedging is to academic English — it’s how writers stay credible while admitting what they don’t know.

The usual suspects:

• Modal verbs: *may, might, could, would, can*
• Approximators: *somewhat, relatively, to some extent, in some cases, generally*
• Distancing frames: *it could be argued that, it seems that, one might suggest, this appears to*
• Softeners: *perhaps, possibly, arguably, tend to, fairly*

Used once or twice with a reason behind them, these make you sound thoughtful. Stacked four to a sentence with no anchor, they make you sound like you’re hiding.

Why hedges and AI text look alike

To understand the flag, you have to understand what a detector measures. It doesn’t read for meaning. It estimates how *predictable* your words are — given everything so far, how surprising is the next word? Common, expected words produce a low-surprise pattern. Rare, specific, slightly-off-script words produce a high-surprise one. (We walk through the math in the statistics behind AI detectors.)

Stock hedges are among the most predictable strings in all of academic English. “It could be argued that” is a frozen template — once the model sees “it could be,” the rest is almost guaranteed. So is “may have a significant impact on.” These phrases are so worn that a detector can practically finish them for you, which is the definition of low surprise.

Now think about what a large language model does when it’s unsure. It hedges. Trained to avoid confident-but-wrong answers, it reaches for “may,” “could,” “in some cases,” “it is important to consider.” The default voice of a careful AI is *wall-to-wall hedging*. So an ESL writer who hedges everything and a language model playing it safe end up producing the same fingerprint: smooth, qualified, low-risk, low-surprise. The detector can’t tell them apart, because on the axis it measures, they’re nearly identical.

Why ESL writers hedge more

This isn’t a coincidence, and it isn’t a flaw. Three forces push non-native writers toward heavier hedging.

Politeness norms travel. In many languages and academic cultures, stating a flat claim sounds arrogant or rude. A writer trained to soften assertions out of respect carries that instinct into English, and it shows up as extra modal verbs and distancing frames.

Being wrong in a second language feels riskier. When you’re not fully confident in your English, a hedge is insurance. “This may possibly indicate” protects you in a way “this shows” doesn’t. The less sure you feel about the language, the more you reach for the safety blanket.

Textbooks hand out templates. English for Academic Purposes courses teach hedging explicitly, often with phrase banks: *it could be argued that, this suggests that, there is some evidence to.* Those phrases are useful — and they get reused until they’re reflexes, which is exactly how you build a predictable, repetitive surface.

Stack all three and you get prose that’s polite, cautious, and statistically hard to tell apart from a hedging machine. This lines up with the 2023 study by Liang and colleagues, which found AI detectors misclassified non-native English essays as AI far more often than native ones. Hedging is one of the mechanisms that can feed that gap.

A mini-scenario

Imagine a grad student writes a literature-review paragraph like this:

> *It could be argued that social media may have a somewhat negative effect on adolescent well-being. This appears to suggest that, in some cases, increased usage might possibly be associated with higher levels of anxiety. However, it is important to consider that the evidence may not be entirely conclusive.*

Every claim is technically true and every hedge is “correct” by textbook rules. A detector will likely score it as heavily AI. Look at why: there’s no specific fact in three sentences, the rhythm is even, and almost every content word sits inside a stock hedge template. It’s maximally safe — which is maximally predictable.

Now picture the same point rewritten with real anchors (the specifics below are illustrative):

> *One large study links heavy social-media use to higher anxiety in teens, though the reported effect was small and self-reported. Later work pointed the same direction but found a weaker correlation. So the link looks real but modest — not the crisis some headlines claim.*

Still cautious. She kept “small,” “modest,” “weaker.” But the caution now points at *specific limits* — effect size, measurement method, mixed replication — instead of floating free. Naming concrete details raises the surprise level. She didn’t become arrogant. She became specific.

How to fix it without faking confidence

You don’t delete hedging. You upgrade it.

Hedge with a reason, not a reflex

Bad: “This may suggest a possible relationship.” Better: “This suggests a relationship, but the sample was small.” A real hedge names *why* you’re cautious. An empty one just sprays softeners. Real hedges read as competent; empty ones read as filler — to detectors and to human graders both.

Commit where you actually have a view

If you believe something, say it in one plain sentence, then qualify the parts that deserve it. “The policy worked in cities but not rural areas” beats “it could be argued that the policy may have had some effect in certain contexts.” One short, direct claim breaks up the wall of qualifiers and adds the burstiness that even, hedged prose lacks.

Count your hedges per paragraph

Read a paragraph and circle every modal, approximator, and distancing frame. More than two or three in a stretch? Cut the weakest. You’ll usually find one carrying the real meaning and the rest just padding it.

Read it aloud

Stacked hedges *sound* evasive when spoken. Your ear catches “it might possibly perhaps be the case that” faster than your eye does. If you trip over the caution, a reader will too — and so will the statistics.

The honest fix is precision, not deletion. For the bigger picture on why second-language writing trips these tools and what else to watch for, browse the rest of our writing on detection. And remember: a flag is a probability about phrasing, not proof of anything. Save your drafts.

Frequently asked questions

Is hedging bad writing, or just bad for AI detectors?

Hedging is good writing when it’s precise — it shows you understand the limits of a claim. The problem is volume and vagueness. Real academic hedging points at something specific: a small sample, a result that only holds in one context. Empty hedging just softens everything with no reason, and that’s the kind detectors and human graders both dislike. The goal isn’t to remove hedges; it’s to make each one earn its place.

Why do ESL writers tend to hedge more than native speakers?

A few habits stack. Many first languages treat blunt claims as rude, so writers carry that politeness into English. Writing in a second language raises the fear of being wrong, and a hedge feels like insurance. And EAP textbooks hand out hedging templates that get reused until they’re reflexes. None of that is a flaw — it just produces the safe, predictable phrasing detectors associate with machine text.

Does heavy hedging really make a detector flag me?

It contributes. Detectors estimate how predictable your wording is, word by word. Stock hedges are some of the most predictable strings in academic English, so a paragraph built from them reads as low-surprise — the same pattern an AI playing it safe produces. Hedging alone won’t decide a score, but combined with repeated safe vocabulary and even sentence rhythm, it pushes you toward the AI side.

If my hedged essay gets flagged, does that prove I cheated?

No. A detector outputs a probability about word patterns, not evidence about who wrote them. OpenAI retired its own AI classifier in July 2023 for low accuracy, and a 2023 study by Liang and colleagues found these tools misclassified non-native English writing as AI far more often than native writing. A flag is a reason to look closer, never a verdict.

How do I cut hedging without sounding arrogant?

Trade vague hedges for specific ones. Instead of “it could be argued that the policy may have some effect,” name the actual limit: “the trial found a small effect, but the sample was tiny.” You’re still cautious — you just anchored the caution to a real constraint. Where you have a view, state it plainly in one sentence, then hedge the parts that genuinely deserve it.

The takeaway

Your hedging habit isn’t a character flaw and it isn’t proof of anything. It’s what happens when good academic training meets second-language caution — and it lands you in the same statistical neighborhood as a careful AI. The fix keeps your honesty intact: anchor each hedge to a real limit, commit where you have a view, and let specifics do the softening. If you want to see which of your sentences read as machine-cautious before you submit, run your draft through PaperBleach and check the heatmap. If it’s a tool you’ll use often, check the plans.