# AI Detectors and Non-Native English Writers · ScriptGrain

> Detectors flag non-native writing at far higher rates, and the mechanism catches neurodivergent and rule-following writers too. What it means and what to do.

Canonical: https://scriptgrain.com/blog/ai-detectors-and-non-native-english-writers

# AI Detectors and Non-Native English Writers

*By Jack Stovell · 2026-08-29 · Research*

Here's the thing about learning English as a second language properly: someone drills you until you write clean, simple, correct sentences. Subject, verb, object. Vary your openers, but not too much. Avoid idioms until you've mastered the basics. It's good pedagogy. It's also, apparently, exactly what an AI detector is trained to flag.

That's the bias. Non-native English writers, taught to write in the tidy, rule-following style that language courses reward, get caught by detection tools at rates well above native speakers. Not because their writing is machine-generated. Because it's careful. And careful, to a detector, looks suspiciously like a template.

Why this happens mechanically

Detectors don't read for meaning. They read for statistical pattern, mostly a measure called perplexity: how predictable is each word, given what came before? Low perplexity means safe, expected word choices. High perplexity means surprising ones, the kind a fluent, idiosyncratic native writer produces without thinking. Detectors flag low perplexity as a machine tell, because language models genuinely do favour the statistically safest next word.

Except that's also what a careful non-native writer produces. You learned to avoid risky phrasing. You learned the safe collocations, the textbook transitions, the reliable sentence shapes that don't trip you up mid-thought. So does the model. Same surface pattern, completely different cause. The detector can't tell the difference, because it was never built to look for cause. Just pattern.

And this is not a fringe finding. It's documented, consistent, and worth reading in full rather than trusting a paraphrase, so go and look at the sourced numbers at scriptgrain.com/reference/ai-detector-accuracy. I'm not going to invent statistics here; the reference page exists precisely so you don't have to take my word for it.

It's not only non-native writers, either. Neurodivergent writers who lean on formulaic structure because it's cognitively reliable get swept up the same way. So do very rule-following writers generally, the ones who learned "good writing" from a style guide and never deviated. The mechanism doesn't care why your prose is tidy. It only registers that it is.

What it means if you're affected

So you've been flagged. Or you're worried you will be. Here's what actually matters: the flag is not evidence. It's a probability score dressed up as a verdict, and probability scores are not proof of authorship.

What is proof, or close to it? Evidence. Concrete, checkable, boring evidence. Version history is the big one: drafts, timestamps, edit trails, the mess of revisions that shows a piece evolving over hours or days. A machine doesn't produce that mess. You do, because you second-guess your third paragraph and rewrite your opening line four times, and all of that leaves a trail.

Consistent personal voice is the other one. If your writing has a stable, measurable pattern across many pieces, your sentence lengths, your habits with commas, your recurring phrases, your particular way of building an argument, that consistency is itself evidence. Not a vibe. A pattern, and patterns can be measured. That's what a stylometric voice profile actually captures: not "does this sound AI", but "does this sound like the same person who wrote these other twelve things". Forty-five separate measured attributes, in ScriptGrain's case, from sentence rhythm to vocabulary diversity to how you punctuate. That's not proof of humanity, either, but it's proof of you, which is the more useful claim anyway.

If you want to see this in action rather than take my word for it, there's a free browser-based tool at scriptgrain.com/tools/writing-style-analysis that measures your own writing style directly, no account, no card. You can also build a full voice profile and run one analysis for free, and plans start from £12 a month if you want it ongoing. None of that changes what you write. It documents what you already write, which is a different job entirely.

The case against changing how you write

Here's where I get stubborn. The instinct, once you know detectors flag tidy prose, is to write messier on purpose. Throw in a typo. Vary your sentences artificially. Add some texture you don't naturally have, to dodge a machine that was wrong about you in the first place.

Don't.

That's not a style choice, it's a capitulation, and it solves the wrong problem. You'd be distorting your actual voice to satisfy a tool with a documented accuracy problem, one that already penalises you unfairly for writing the way you were taught to write, or the way your brain finds workable. Bending your prose to flatter a flawed detector doesn't fix the detector. It just adds a second layer of distortion on top of the first.

The fix isn't performing imperfection. It's evidence. Keep your drafts. Keep your history. Let your genuinely consistent voice, whatever shape it takes, stand as its own record across everything you write. That's a stronger defence than any amount of stylistic contortion, because it's true, and true things hold up under scrutiny in a way performed quirks never do.

At the end of the day, a detector's flag is an opinion generated by a pattern-matcher with known blind spots, particularly around exactly the writers this piece is about. It's not a verdict on your authenticity. Your version history is. Your consistent, distinctive voice across time is. Measure it if you want proof; don't mutilate it to dodge a guess.

Fair enough if that sounds like a lot of faith to place in "just keep writing normally". It is. But it's also the only approach that doesn't ask you to become a worse, more anxious version of yourself to satisfy software that gets this wrong more often than it should.

[More from the ScriptGrain Journal](https://scriptgrain.com/blog)