French Speakers in English Academia: Cognates, Formality, and the AI Flag Problem

You wrote the essay yourself. Every sentence. Then the detector hands back a number that says “92% likely AI,” and your stomach drops. If French is your first language, this happens more than it should, and it isn’t because your English is bad. It’s because your English is *good* in a very particular way that detectors don’t know how to read.

Key takeaways

• AI detectors score how statistically predictable your wording is. French speakers often pick the predictable, formal option, which reads as machine-like.
• French and English share thousands of cognates, so French writers reach for the Latin-rooted word (“utilize,” “demonstrate,” “commence”) over the plain one (“use,” “show,” “start”).
• Academic French rewards smooth, impersonal, tightly-structured prose. Carried into English, that polish overlaps with the patterns detectors tie to AI.
• A high score is a probability, not proof. Detectors show documented bias against non-native writers, and the false-positive rates some tools publish are disputed and far too uncertain to hang an accusation on.
• You don’t fix this by writing worse English. You fix it by varying rhythm, mixing in plain words, and letting your voice through.

What the detector is actually reacting to

Here’s the thing most people get wrong: a detector doesn’t “know” anything about French. It can’t tell that you grew up in Lyon or Montreal. What it does is measure how surprising your next word is, statistically, given the words before it. Text where every word is the most expected choice scores as machine-like. Text with unexpected turns scores as human.

That single mechanic is why French speakers get caught. Once you see that detectors are mostly measuring predictability — not creativity, not intelligence, not whether you cheated — almost everything that follows makes sense.

The cognate trap

French and English share an enormous vocabulary, mostly through Latin and Norman French. That sounds like an advantage, and for comprehension it is. For sounding human on a detector, it’s a quiet liability.

When a French speaker reaches for an English word, the cognate is right there, and it usually *is* a real English word. The problem is which register it sits in. In English, the Latin-rooted cognate is almost always the formal one, and the plain Anglo-Saxon word is the everyday one:

• *utilize* vs. use
• *demonstrate* vs. show
• *commence* vs. start
• *facilitate* vs. help
• *sufficient* vs. enough
• *obtain* vs. get
• *necessitate* vs. need

A native English speaker bounces between both columns without thinking. A French speaker, pulled toward the familiar cognate, drifts toward the left column over and over. Each individual choice is correct. Stacked across 1,500 words, they build a uniformly high, formal texture — and “formal and uniform” is exactly the fingerprint detectors associate with generated text. The model finds your wording predictable because you keep picking the option it would have picked.

Formality isn’t politeness here — it’s a pattern

French academic writing has its own conventions, and they’re good ones. Argumentation tends to be explicit and well-signposted. The tone is impersonal; you efface yourself from the text. Sentences get built carefully, often long and balanced, with subordinate clauses nested cleanly inside the main one. The result reads as serious, controlled, professional.

Now look at what AI-generated English tends to look like: impersonal, evenly structured, every sentence roughly the same comfortable length, transitions placed like clockwork. The overlap is uncomfortable. The habits that earn top marks in a French *dissertation* produce, in English, a rhythm a detector reads as suspiciously smooth.

The usual word for the missing ingredient is *burstiness* — the natural unevenness of human writing, where a long sentence is followed by a short one. Three words. Then a sprawling clause that wanders off somewhere before circling back. Native writers do this without noticing. Trained-formal writing, French or otherwise, sands it flat.

A mini-scenario: Camille’s literature review

Camille is a French master’s student writing a literature review in English. Her draft opens:

> “This paper will demonstrate that social media utilization facilitates political engagement. Numerous studies have established a correlation between platform usage and participation. It is necessary to examine these findings in order to comprehend the underlying mechanisms.”

Every word is real, the grammar is clean, the argument is sound. A detector returns 89% AI. (That number is made up to make the point — but the pattern behind it isn’t.) Why so high? *Demonstrate, utilization, facilitates, established, comprehend, in order to* — six formal Latinate choices in three sentences, all the predictable option, all in sentences of nearly identical length and shape. Low surprise, no rhythm.

Here’s a revision that says the same thing:

> “Social media helps people get politically involved — that’s the case I want to make here. Plenty of studies already link platform use to participation. But the numbers alone don’t tell us *why*, and that’s the part worth digging into.”

Notice what changed. Plain words sit next to formal ones. Sentence lengths jump around. A dash, an italic, a first-person “I.” The meaning is identical; the texture is human. This isn’t worse academic writing — plenty of supervisors would call it clearer. It just stopped being uniform.

How to keep your English strong and stop sounding like a machine

You’re not trying to break your English. You’re adding back the variation that formal training trained out of you. A few concrete moves:

Mix your vocabulary columns on purpose

Don’t ban cognates — that would be silly, and some are the precise word you need. Just notice when you’ve used the formal option three times in a paragraph and swap one for the plain English word. “Use” instead of “utilize.” “Show” instead of “demonstrate.” Aim for a blend, not a single register.

Vary sentence length deliberately

After a long, clause-heavy sentence, write a short one. Five words is fine. The contrast is the point. If you read your paragraph aloud and it has a steady, even drumbeat, that’s the sound a detector flags.

Let yourself into the text

French academic style trains you to disappear. English academic style, especially in the humanities and social sciences, often tolerates more presence. A “here’s the case I want to make” or a “this surprised me” adds the unpredictability — and the actual humanity — detectors can’t fake.

Don’t translate, then polish — draft in English

When you draft mentally in French and render it into English, you import French sentence structure and grab cognates by reflex. Drafting straight in English, even if it feels rougher at first, gives you more native rhythm. The roughness is a feature here.

If you’ve already got a stiff draft and want to loosen it without losing your argument, a humanizer can vary the rhythm for you — but treat it as a finishing pass on writing you stand behind, not a way to disguise something you didn’t write.

When you’re flagged anyway: it’s a probability, not a verdict

Even careful writers get flagged. So it helps to know what the number is and isn’t.

A detector score is a probability estimate. It is not evidence. OpenAI quietly retired its own AI Text Classifier in July 2023 because its accuracy was too low to be useful — and that was the company that built the models everyone’s trying to detect. A 2023 Stanford study (Liang et al., published in *Patterns*) found several detectors were biased against non-native English writers: more than half of the human-written TOEFL essays they tested got misflagged as AI, while essays by native-English US students were classified almost perfectly. The authors point to low perplexity — exactly the predictability we’ve been talking about — as the likely cause. If you’re a French speaker who got a high score, you’re not imagining the pattern. It’s documented.

Practically, that means two things. First, keep your process. Version history in your word processor, your notes, your outline, your reading — that record is worth more than any score because it shows the work. Second, if you’re ever asked to explain a flag, you can. You can point to your drafts, and you can explain that your formal vocabulary and tidy sentences come from how you were taught to write, not from a chatbot. If you want a fuller game plan, here’s how students can push back when a detector gets it wrong. A number can’t argue back. You can.

Frequently asked questions

Why would a detector flag my English just because French is my first language? It isn’t flagging your French. It reacts to patterns in your English that are common among French speakers — formal Latin-rooted vocabulary, long even sentences, an impersonal tone — and those make your text more predictable. Predictability is what detectors score. The 2023 Stanford study found related bias against non-native writers.

Should I write worse English to avoid being flagged? No. The goal is more *varied* writing, not worse writing. Mix sentence lengths, use the plain word when it fits, and let some personal phrasing through. Natural English actually scores better than uniformly formal English.

Are cognates always a problem? No. They’re often the right word. The issue is over-reliance — reaching for “demonstrate” every time instead of sometimes saying “show” builds the uniform, high-register texture detectors associate with generated text.

Can I prove I wrote it myself if I’m flagged? Often, yes. Keep your version history, outlines, notes, and sources. A score is a probability, not evidence. If you can show your process, you’re in a far stronger position than the tool’s number on its own.

Does this apply to professional writing too? Yes. Reports and web copy by French speakers carry the same texture. The fixes are identical: vary rhythm, prefer plain words where they read naturally, add concrete detail.

The short version

Your English isn’t the problem. The *evenness* is. French training pushes you toward formal cognates and smooth, balanced sentences, and detectors mistake that polish for a machine. Loosen the rhythm, mix in plain words, and let your voice show — that’s better writing on its own terms, and it happens to read as human.

Want to see how your draft scores before you submit it, then fix only the parts that read flat? Try PaperBleach. And if you want more on writing and detection, the blog goes deeper.