A student submits an essay they spent two weeks writing. Their professor runs it through an AI detector and it comes back 85% AI-generated. The student had never used ChatGPT. They wrote every word themselves, in their second language, with more care and effort than most native speakers bring to the same task.
This situation is happening in universities, companies, and publishing houses around the world — and it is happening far more often to people writing in their second or third language than to native English speakers. It is not a minor edge case. It is a systematic flaw in how AI detection currently works, and anyone who uses these tools should understand it.
This article explains the exact technical reason it happens, who is most affected, and what to do about it — whether you are the writer being flagged or the person making the decision.
The real scope of the problem
The false positive problem with non-native writers is not anecdotal. Multiple researchers and educators have documented it. A 2023 study by researchers at Stanford found that GPTZero and other detectors flagged TOEFL essays — written by real humans — as AI-generated at significantly higher rates than essays written by native English speakers. Some detectors flagged genuine human ESL writing as AI more than 60% of the time.
The situation has improved slightly as detectors have been updated, but the underlying technical problem has not been solved. The gap between how native and non-native English writing scores on detection metrics is still substantial.
When a student is accused of AI use based on a detector score, the consequences can include failing grades, academic misconduct proceedings, and lasting reputational damage. The burden of proof shifts unfairly to the writer. And the people least equipped to defend themselves — those with the least institutional power, often international students — are the ones most likely to be falsely flagged.
The technical reason it happens
To understand this, you need to know what AI detectors actually measure. Most detectors use two core signals: perplexity and burstiness.
How predictable each word choice is. Low perplexity means the words are statistically common and expected. High perplexity means the choices are surprising or unconventional. AI writing has low perplexity because models select the most probable next word. Human writing — especially native, informal, idiomatic writing — has higher perplexity.
How much sentence lengths vary. High burstiness means long and short sentences alternate naturally. Low burstiness means all sentences arrive at similar lengths — the flat, metronomic rhythm characteristic of AI output.
Now think about how a careful non-native English writer typically writes. They choose common, familiar vocabulary rather than taking risks with idiomatic or unusual words. They write grammatically correct, complete sentences at consistent lengths rather than using fragments or very long complex clauses. They avoid contractions and colloquialisms because those feel informal and unsafe in a language they are still mastering.
Every single one of those sensible, careful habits lowers perplexity and burstiness scores — the same two metrics that detectors use to flag AI text. The overlap is not a coincidence. It is a structural collision between two different things that look identical to a statistical classifier.
We covered perplexity and burstiness in technical detail in our previous post: What Is AI Perplexity and Burstiness? — that article explains the mechanics if you want the full picture.
Who is most affected
International students
Writing academic essays in English as a second language while following formal style guides — exactly the conditions that produce low perplexity and flat burstiness.
Professional ESL writers
Freelancers and content writers working in English who default to safe, common vocabulary rather than idiomatic phrasing they are less confident about.
Academic researchers
Scientists and researchers writing in English — whether native speakers or not — whose field requires formal, structured, passive-voice writing that scores low on both metrics.
Writers from high-context cultures
Writers from cultures where formal, restrained prose is the expected style often produce text that looks statistically flat to English-trained detectors.
The specific writing habits that trigger detectors
These are not bad writing habits. They are rational, careful choices that make total sense when writing in a second language. The problem is not the writer — it is the detector treating these patterns as evidence of AI authorship.
| Writing habit | Why ESL writers do it | Why detectors flag it |
|---|---|---|
| Choosing common, familiar vocabulary | Reduces risk of using a word incorrectly in an unfamiliar language | Common words are highly predictable — this lowers perplexity directly |
| Consistent sentence lengths | Easier to control grammar and structure in a second language | Uniform sentence lengths signal low burstiness — an AI pattern |
| Avoiding contractions | Contractions feel informal; formal writing feels safer and more correct | AI also avoids contractions in formal contexts — another signal overlap |
| Structured, predictable paragraph format | Following the paragraph structures taught in language courses | Topic sentence → explanation → example is exactly the AI paragraph template |
| Formal transition phrases | Learned as correct academic connectors in language training | "Therefore," "As a result," "In addition" are classic AI transition markers |
| No personal anecdotes or opinions | Academic writing conventions in many countries avoid first-person perspective | Absence of personal voice is a strong AI signal for most detectors |
Reading this list, the injustice becomes clear. A student following exactly the writing instructions they were given — be formal, be structured, use academic transitions — ends up producing text that looks maximally AI-like to a detector.
What to do immediately if your writing is flagged
If you are a writer whose genuinely human work has been flagged, here is a practical response framework.
-
1Do not accept the verdict silently
A detector score is not evidence. It is a statistical estimate with a known false positive problem. You have every right to contest it. Accepting it without pushing back sets a precedent that harms you and everyone else in the same situation.
-
2Gather your writing process evidence
Collect anything that shows the work is yours: draft versions saved with different timestamps, browser history showing research, notes taken while writing, an outline created before the final version, or citations you personally tracked down. These show a process that AI does not have.
-
3Run the text through multiple detectors yourself
Different detectors produce different results on the same text. If Humanify's AI Detector gives a different result than the tool used to flag you, that disagreement itself is evidence of the tool's unreliability on your specific writing style. Screenshot all results.
-
4Request a live writing demonstration if possible
In academic contexts, offering to write a short piece on the same topic in front of the assessor is a powerful way to demonstrate authorship. Your writing style, vocabulary choices, and knowledge will be consistent with the flagged piece.
-
5Reference the published research on false positives
Research documenting high false positive rates for non-native English writers is published and credible. Citing it in a formal appeal changes the conversation from "I say I didn't use AI" to "the tool itself is known to produce wrong results on writing like mine."
How to adjust your writing without losing your voice
This section is for writers who want to reduce false positive risk going forward — not by writing worse or less carefully, but by adding a few specific elements that naturally signal human authorship to detectors.
The key principle: you do not need to change your vocabulary, your formality level, or your overall approach. What you need to add is specificity and rhythm variation. Both are compatible with careful, formal writing.
Add specificity to raise perplexity
Research has shown that regular physical exercise has a positive impact on mental health outcomes. Studies indicate that individuals who exercise frequently report lower levels of stress.
A 2021 meta-analysis of 23 randomised controlled trials found that aerobic exercise reduced symptoms of depression by approximately 48% compared to control groups — a result that held across age groups and exercise intensity levels.
You do not need personal anecdotes to raise perplexity. Academic writing can be specific through citations, exact statistics, study details, and named researchers. These are already expected in academic writing — using them more precisely helps your detector score without changing your style.
Add rhythm variation to raise burstiness
This is easier than it sounds in formal writing. You do not need to write casually — you just need occasional contrast in sentence length. Long academic sentence establishing a complex idea, supported with evidence and nuance across multiple clauses — then one short sentence that states the conclusion plainly.
The results demonstrated a significant correlation between sleep duration and academic performance. Students who slept fewer than six hours showed measurably lower test scores. The relationship remained consistent across different age groups and subjects.
Students who slept fewer than six hours showed measurably lower test scores across every subject and age group tested, with the correlation remaining consistent even after controlling for socioeconomic factors, prior academic performance, and attendance rates. Sleep matters.
Two-word or three-word sentences are perfectly acceptable in formal writing when used deliberately. "The results held." "This is significant." "One exception exists." They are not informal — they are emphatic. And they do the most work of any technique for raising burstiness quickly.
The full checklist for non-native writers
- Replace at least one general claim per section with a specific statistic, date, study name, or named researcher
- Include at least one sentence under 10 words in every major paragraph
- Vary paragraph length — not every paragraph needs to be 4–5 sentences
- Where your field permits it, include one direct observation or qualified opinion per section
- Replace "it has been shown that" with the actual study or source where possible
- Do not change your vocabulary to seem more informal — that is not what detectors are measuring
- Do not add fake personal anecdotes — specificity from real sources works equally well
- Do not run your text through a paraphrase tool — it will not fix the structural problem and often makes it worse
Run your text through our AI Detector to see exactly which sentences are triggering the flag — then use the AI Humanizer to adjust rhythm and structure.
Check Your Text FreeFor educators and employers: how to handle this fairly
If you are making decisions based on AI detector results, these points are worth taking seriously.
A detector score is not evidence of AI use. It is a probabilistic estimate trained on a dataset that underrepresents non-native English writing. Using it as the primary basis for an academic misconduct finding or an employment decision is applying a tool far outside its validated use case.
Several things a detector score cannot tell you:
- Whether the writer used AI as a drafting tool and then substantially rewrote the output
- Whether the writing style is genuinely the writer's own, developed over years of careful formal writing
- Whether the writer is a non-native speaker whose careful English happens to look like AI output statistically
- Whether the result would be different on a different detector
A detector score should prompt a conversation, not a verdict. Ask the writer to explain their process. Ask them to write something short in your presence. Ask for draft evidence. These approaches give the writer an opportunity to demonstrate authorship that a statistical tool cannot fairly evaluate.
The institutions that have handled this best are those that treat AI detector results the way they treat plagiarism similarity scores from tools like Turnitin — as a flag that warrants investigation, not as proof of wrongdoing. A 70% similarity score on Turnitin does not mean a student plagiarised. A 70% AI score does not mean a student used AI. Both require human judgment on top of the tool output.
Tools that help non-native writers in this situation
If you need to reduce your false positive risk, two tools on this site are directly relevant:
Humanify's AI Detector shows you sentence-level analysis — which specific sentences are triggering the flag. This is more useful than an overall score because it shows you exactly where to focus your edits, rather than leaving you guessing about the whole document.
Once you know which sections are flagging, Humanify's AI Humanizer can help reprocess those specific sections to vary rhythm and reduce surface-level pattern matches — without requiring you to rewrite them manually. It handles the burstiness adjustments automatically, which is the most mechanical part of the fix.
The combination — detect which sentences are flagging, then adjust rhythm in those specific sections — is more efficient than trying to rewrite an entire document and more targeted than applying a blanket paraphrase tool to everything.
For managing the word count and length of your revised sections, the Word Counter gives you a quick read on document length and average sentence metrics without needing to paste into a full analysis tool.