Code-Switching in Writing: When Mixing Formal and Casual English Looks Like AI

You write the way you think, and if you grew up moving between two languages, your English probably moves too. One paragraph reads like a textbook; the next sounds like you’re talking to a friend. That range is a skill. But to an AI detector, which sees patterns and not people, those shifts can look less like a bilingual writer flexing and more like something stitched together by a machine.

This is one of the quieter reasons ESL and bilingual students get false-flagged. It isn’t about grammar mistakes. It’s about texture.

Key takeaways

• Code-switching – shifting between formal and casual English in one piece – is normal for bilingual and ESL writers, but it can create the uneven texture detectors misread.
• Detectors don’t understand “register.” They measure statistical patterns like predictability and sentence variety, so a jump from textbook-formal to chatty can read as two writers, or as a machine.
• Over-correcting toward one flat, formal tone is the trap: perfectly uniform prose actually looks *more* AI-like.
• The fix isn’t to erase your range. It’s to make register shifts intentional and signposted, so they feel earned.
• If a draft reads choppy, check it against a detector first, then smooth the seams instead of flattening your whole voice.

What code-switching actually means here

In linguistics, code-switching usually means alternating between languages – dropping a Spanish phrase into an English sentence, say. But there’s a quieter version that happens entirely within English: register-switching. You move between formal and casual modes depending on context, audience, and habit.

Most people do this. The difference for bilingual and ESL writers is that the *boundaries* between registers can be sharper. You might have learned formal academic English in a classroom and casual English from movies, friends, or the internet – two separate streams that didn’t fully blend. So when you write, the seams between them show more than they would for someone who absorbed both at once.

That’s not a flaw. But it does create a specific kind of unevenness on the page.

Why detectors misread register shifts

AI detectors don’t read for meaning. They run statistics. Two of the big ones are how *predictable* your text is (often called perplexity) and how much your sentences *vary* in length and rhythm (burstiness). Human writing tends to be a little unpredictable and pretty bursty – long sentences next to short ones, surprising word choices, the occasional fragment. Machine text, especially older models, tends to be smooth, even, and predictable. If you want the deeper version of how those numbers work, we wrote up the statistics behind AI detectors separately.

Here’s where code-switching gets caught in the gears. When you shift registers abruptly, you sometimes produce text that’s locally *very* predictable – a stretch of stiff, formulaic academic phrasing – and then snap to something casual. To a sentence-level detector scanning chunk by chunk, that formal stretch can score high for “AI-like” on its own. The casual part might score human. The piece as a whole looks inconsistent, like it had more than one author. And inconsistency lowers the model’s confidence that a single human sat down and wrote the thing.

Detectors output a probability, not proof. But when a teacher sees a paragraph flagged at 80%, that nuance often gets lost.

A mini-scenario: Mariam’s literature essay

Mariam is a sophomore. Spanish is her first language; she’s strong in English but learned it in two modes. Here’s a stretch from her essay on *The Great Gatsby*:

> The novel utilizes the green light as a multifaceted symbol that represents the unattainable nature of the American Dream. Furthermore, Fitzgerald employs this motif to underscore the protagonist’s perpetual yearning. Honestly though, Gatsby is kind of just a guy who can’t let go of his ex, and you feel bad for him because he tried so hard for nothing.

Read that out loud. The first two sentences are textbook-formal – the kind of phrasing you reach for when you’re nervous and trying to sound “academic.” Then she drops into how she’d actually explain it to a classmate. Both halves are hers. But the cut between them is jarring, and the formal half is exactly the kind of predictable, synonym-heavy prose (“utilizes,” “multifaceted,” “perpetual yearning”) that detectors associate with machine generation.

A detector might flag the first two sentences and clear the third. Mariam didn’t cheat. She just hasn’t blended her two registers, and the abrupt seam made the formal part look generated.

The trap: don’t flatten everything

The instinct after reading something like that is to “fix” it by making the whole essay formal. Cut the casual aside, make every sentence sound like the first two. Don’t.

Uniform, over-polished, relentlessly formal prose is itself one of the most reliable AI tells. Real human writing breathes. It has range. If you sand every sentence down to the same flat academic register, you trade one detection risk for another – and you lose the part of your voice that made the writing yours. We’ve all read the essay that sounds like it was written by a committee of robots. That’s the failure mode on the other side.

The answer isn’t less range. It’s better-managed range.

How to make register shifts work for you

Signpost the switch

The casual aside in Mariam’s paragraph isn’t the problem; the *lack of a bridge* is. A short transitional beat prepares the reader. Compare:

> Fitzgerald uses the green light to represent a dream Gatsby can never reach. Strip away the symbolism, though, and he’s just a man who can’t let go of someone – and that’s what makes him hard to dismiss.

Same two registers. But “strip away the symbolism, though” tells the reader you’re shifting on purpose. The seam is now a doorway instead of a crack.

Vary sentence length on both sides of the shift

If your formal sections are all long and your casual sections are all short, the divide gets louder. Slip a short sentence into the formal part and let a casual thought run long sometimes. The mixing softens the contrast and, conveniently, boosts the natural variety detectors read as human.

Watch the “thesaurus formal” reflex

A lot of register-jump trouble comes from over-formalizing. “Utilizes” instead of “uses.” “In order to facilitate” instead of “to help.” That inflated diction reads as both unnatural and machine-ish. When you catch yourself reaching for the fancy synonym, ask whether the plain word is actually fine. Usually it is.

Read it aloud

This is the cheapest tool you have. When you read your draft out loud, the spots where you stumble or wince are almost always the unmanaged seams – the place where the formal mask slips off and the casual voice barges in with no warning. Mark those. Those are the sentences to smooth, not the whole essay.

Check, then smooth – don’t pre-flatten

If you’ve revised a lot and you know your draft swings between registers, a sentence-level scan can show you which seams read as off, so you can fix those specific transitions instead of bleaching your whole voice into one tone. There’s a lot more on detection and writing if you want to go deeper into how the tools behave, and if you’re weighing whether a tool is worth it, what PaperBleach costs is laid out plainly.

The bias underneath all this

It’s worth naming why this hits bilingual writers harder. A 2023 Stanford study by Weixin Liang, James Zou, and colleagues found that several AI detectors disproportionately flagged essays written by non-native English speakers – across a set of TOEFL essays, more than half were misclassified as AI, with a false-positive rate above 60%. Part of the reason was that this writing tended to use more limited, more predictable vocabulary, which is exactly what detectors are tuned to suspect. (For what it’s worth, OpenAI quietly retired its own AI text classifier in July 2023, citing low accuracy. Even the people building these tools know they’re imperfect.)

So if you’re bilingual, you’re already starting from a slightly unfair baseline. Unsignposted register shifts stack risk on top of that. None of this means your writing is the problem – the tools are biased, and that’s documented. But understanding the mechanism lets you write in a way that’s both more natural *and* less likely to get caught in a bad detector’s net.

Frequently asked questions

Does code-switching itself get flagged as AI? Not directly. Detectors have no “code-switching” signal. They measure predictability and sentence variety. The trouble is that abrupt, unmotivated register shifts produce jumpy texture that lowers a detector’s confidence one human wrote the piece.

I’m bilingual and shift registers naturally. Should I stop? No. Flattening your range into one stiff tone usually backfires, since uniform over-formal prose is a common AI tell. Aim for intentional switching: signpost shifts, smooth transitions, and read aloud to catch the seams.

Why do non-native English writers get flagged more often? The 2023 Stanford study found detectors disproportionately flag non-native writing, partly because it uses more predictable vocabulary – the very thing detectors associate with machines. The bias is in the tools, not your writing.

How do I smooth a register shift without sounding robotic? Add a short transitional phrase, vary sentence length around the shift, and don’t drop slang straight after textbook phrasing with no bridge. Make the change read as a choice, not an accident.

Should I run my essay through a detector before submitting? It’s worth it if you’ve revised heavily or your draft swings between registers. A detector gives a probability, not proof, but it points to the exact sentences that read as off.

Closing

Your ability to move between formal and casual English is a real skill, not a liability. The fix for register-jump false flags was never to erase that range – it’s to make your shifts deliberate, signposted, and smooth. Write in your full voice, then tighten the seams. If you want to see which sentences read as off before you submit, run your draft through PaperBleach and smooth the parts that need it.