Portuguese and Brazilian Students: Why Your English Drafts Read as AI-Generated

You wrote the whole thing yourself. In English. No ChatGPT, no shortcuts. And the detector still lights up red: “88% likely AI.” If Portuguese is your first language, whether you grew up in São Paulo, Lisbon, or Recife, this happens far more than it should. It’s not because you cheated. It’s because the way many Portuguese speakers naturally write English overlaps, by pure accident, with the exact patterns these tools were trained to catch.

Key takeaways

• AI detectors score statistical patterns, not who wrote the text. A human-written essay can still read as machine-made.
• Portuguese-to-English habits, false friends, gerund-heavy phrasing, and formal Latinate vocabulary, happen to land in the patterns detectors flag.
• This is a documented bias, not a personal failing: a 2023 Stanford study found detectors flag non-native English writers far more often than native speakers.
• A detector returns a probability, never proof. A high score is a reason for a human to look closer, not a verdict.
• You can lower false flags by varying sentence length, cutting gerunds, and choosing plain words, all without losing your voice.

What the detector is actually measuring

Here’s the part most people miss. An AI detector doesn’t read your essay the way a professor does. It doesn’t weigh your argument or judge your honesty. It runs a math check on the surface of the words.

Two signals do most of the work: perplexity and burstiness. Perplexity asks how surprising each word is, given the words that came before it. Low perplexity means the text is very predictable, the “obvious next word” flow that language models produce. Burstiness asks how much your sentence length and rhythm jump around. Real people write in bursts, a long, winding sentence followed by a blunt three-word one. Machines tend to hum along at a single steady pace. If you want the full breakdown, we wrote a deeper piece on the statistics behind AI detectors.

So the tool isn’t really asking “did a human write this?” It’s asking “does this look statistically like the text machines produce?” Those are two different questions. And the gap between them is exactly where Portuguese speakers get caught.

Why Portuguese-influenced English reads as “machine-like”

Nothing below is a flaw in your writing. These are normal, often careful choices. They just happen to sit in the danger zone.

1. False friends and over-literal phrasing

Portuguese and English share a deep Latin root system, which is usually a gift. It also breeds *falsos cognatos*, words that look identical but drift in meaning, plus phrases that come out translated word for word.

You write “pretend” when you mean *pretender* (to intend). “Actually” creeps in where *atualmente* means “currently.” “Push the date” instead of “move the date,” echoing *empurrar*. “Make a question” instead of “ask a question,” straight from *fazer uma pergunta*. None of this is wrong enough to break the sentence. That’s the problem. It produces grammar that’s clean but statistically narrow, tidy, low-surprise constructions the detector reads as predictable. A native speaker’s messier idiom would, ironically, score as *more* human.

2. The gerund habit

Portuguese, especially Brazilian Portuguese, loves the present continuous. *Estou estudando, estamos trabalhando, vai ficando.* That comfort carries straight into English, so drafts fill up with “-ing” forms: “By analyzing the data, we are seeing that companies are improving their results by focusing on customers.”

Stack enough gerunds and the prose gets a smooth, gliding texture, every clause flowing into the next at the same gentle slope. That low burstiness is precisely what the rhythm check punishes. Plain past and present verbs, “we analyzed,” “companies improved,” break the glide and read more human.

3. Formal, Latinate vocabulary

Academic Portuguese rewards an elevated register. So words that feel natural and respectful to you, “utilize,” “demonstrate,” “evidence” (as a verb), “posteriorly,” “realize” in the sense of *realizar* (to carry out), pile up in one paragraph. To you that’s just good writing. To a detector it’s the same polished, elevated surface that language models default to. The score climbs.

Plain English does the opposite. “Use,” “show,” “later,” “do” are shorter, blunter, and harder for the model to predict. Less elegant on paper, more human to the math.

4. Connector scaffolding

Portuguese composition teaches explicit logical signposting. *Portanto, entretanto, além disso, dessa forma, em contrapartida.* Translated, you get a paragraph guarded by “therefore,” “however,” “moreover,” “in this way,” “on the other hand.” Clear, organized, exactly what your teachers wanted. It’s also one of the loudest AI tells there is, because language models lean on those same connectors as glue.

A quick before-and-after

Picture Beatriz, an exchange student from Belo Horizonte writing a discussion post for her business class. Her first draft:

> “Currently, it is fundamental to utilize effective strategies. Moreover, companies are demonstrating value by focusing on their clients. Therefore, brand consistency is becoming essential, since it is facilitating long-term loyalty.”

Every word is hers. She never opened an AI tool. But look at it: false-friend “currently” and “fundamental,” three gerunds, formal Latinate verbs, a connector heading nearly every clause, and four sentences of almost identical length. A detector flags it hard.

Now the same ideas, loosened up:

> “Good strategy matters more than ever right now. Companies that show clients real value, instead of just claiming it, win. That’s where brand consistency comes in. Keep it steady and people stick around.”

Same argument. Same student. But the rhythm jumps, the vocabulary drops to plain words, the gerunds are gone, and the connectors mostly disappear. The second version reads as human because it’s messier in the ways humans are messy. Beatriz didn’t lie about a single thing. She just stopped writing like a textbook.

This bias is documented, not imagined

If this feels unfair, it is, and researchers have actually measured it. A 2023 study out of Stanford, by Weixin Liang and colleagues in James Zou’s lab, ran several popular AI detectors against essays written by non-native English speakers. Those essays were flagged as AI-generated far more often than essays by native speakers. The authors tied it to a fairly mechanical cause: non-native writers tend to use more constrained, less varied language, and detectors read that limited variety as a sign of AI.

In plain terms, the tools penalize a normal stage of learning a language. That’s worth saying out loud if you’re ever sitting across from a professor. You’re not making an excuse, you’re describing a known limitation of the technology.

It’s also worth remembering how shaky these tools can be. OpenAI, the company that builds ChatGPT, pulled its own AI-detection classifier in July 2023 and cited low accuracy. If the maker of the model couldn’t reliably catch that model’s output, a third-party tool flagging your hand-written essay deserves real skepticism. There’s more on how detection actually works on the blog if you want the longer version.

A score is a probability, not a verdict

Schools and writers both forget this one. An honest detector hands back a likelihood, a percentage, a confidence band. Not a confession. “88% AI” means the math leaned that way, nothing more. The tool can’t see your drafts, your sources, or the four hours you spent rewriting the introduction in a café with bad Wi-Fi.

So a flag is the start of a conversation, not the end of one. Your strongest evidence is everything the detector can’t measure: your version history, your outline, your messy notes in Portuguese, and your ability to explain your own argument out loud. A percentage doesn’t survive contact with a writer who clearly knows their material.

How to write so you read as human, without faking a voice

The goal is not to sound like you were born in Florida. It’s to sound like a specific person thinking, which you already are. Keep your meaning. Keep your voice. Change only the surface patterns that confuse the math.

• Vary sentence length on purpose. Write one long sentence that develops a thought, then drop a short, blunt one right after. That contrast lifts burstiness faster than almost anything else.
• Cut the gerunds. When you catch “we are seeing,” try “we saw” or “we see.” The flatter verb breaks the glide that flags you.
• Trust plain words. When you reach for “utilize,” ask whether “use” works. It almost always does, and it scores more human.
• Break the connector chain. You don’t need “moreover” and “therefore” standing guard over every sentence. Let some ideas sit next to each other without a label.
• Read it aloud, in English. Translated phrasing and false friends reveal themselves the moment you hear them. If a line sounds like a phrasebook, rewrite it the way you’d actually say it to a classmate.
• Add something only you would write. A specific example from your own city, a small opinion, a concrete number. Specificity is the hardest thing for a model to fake and the easiest thing for you to add.

Want a second opinion before you submit? You can run your draft through PaperBleach and see which sentences look risky, so a surprise flag never blindsides you in front of a professor or an editor.

Frequently asked questions

Why does my essay get flagged as AI when I wrote every word myself?

Detectors don’t know who wrote your text. They estimate how predictable each word is compared to how a language model would write, then add up how much your sentence rhythm varies. If your English leans on smooth, formal phrasing and even-paced sentences, the math can read as machine-like even though you did all the work. For Portuguese speakers, several natural translation habits push the score up at the same time.

Is it actually true that detectors are biased against non-native English writers?

Yes, and it’s been measured. A 2023 study from Weixin Liang and colleagues in James Zou’s Stanford lab ran popular detectors against essays by non-native English speakers and found they were flagged as AI far more often than native-speaker essays. The authors link it to non-native writing using more constrained, less varied language. That’s a known limitation of the tools, not evidence about any one writer.

Does drafting in Portuguese and translating with DeepL or ChatGPT make a flag more likely?

It can. Machine translation produces statistically smooth, predictable English, which is exactly the texture detectors associate with AI. If you write your ideas in Portuguese and run them through a translator, the English output may score higher than if you’d written the same ideas directly in English, even though the thinking is entirely yours.

If I get flagged, does that prove I cheated?

No. A detector reports a probability, not proof. The score is a prompt for a person to look closer, nothing more. Your drafts, your version history, your notes in Portuguese, and your ability to explain your own argument out loud all carry far more weight than any single percentage.

How do I write so I read as human without pretending to be a native speaker?

Keep your meaning and your voice; only change the surface patterns that confuse the math. Vary your sentence length on purpose, trade gerunds for plain verbs, cut the chain of connectors, and read the draft aloud to catch translated phrasing. The goal isn’t to sound like you grew up in Texas, it’s to sound like a specific person thinking, which you are.

The bottom line

Being bilingual is a strength, not a liability, and it shouldn’t cost you a fair read on your own work. The patterns that trip detectors are surface habits, not proof of anything, and you can adjust them without losing a word of what you actually mean. If you want to check drafts regularly, take a look at PaperBleach pricing, and keep this in mind: a percentage is the start of the conversation, not the end of it.