AI Detectors Penalise Non Native Writers: What to Document

AI detectors penalise non native writers because the features they measure — low sentence-length variation, common transition phrases, plain vocabulary — overlap with what careful second-language writing looks like. That overlap produces false positives. It is a documented pattern, it is not proof you cheated, and there are concrete steps you can take: understand the signal, keep a drafting trail, raise it calmly with your institution, and edit your own prose instead of flattening it.

Key takeaways

• Detector output is a probability estimate about text patterns, not a finding about authorship.
• Non-native writers are over-flagged because careful, textbook-shaped English resembles model output on the features detectors measure.
• You can raise a false positive without sounding defensive: ask what evidence standard applies, then supply your own.
• Keep your drafts, version history and source notes — process evidence is the strongest thing you own.
• Edit for rhythm and specificity, not for detector approval; smoothing your syntax to please a tool costs you your voice.

Why does an AI detector flag clean non-native English?

Most detectors do not look for a watermark or a hidden signature. They look at statistical properties of your sentences and compare them to patterns seen in model output. The usual signals are sentence-length uniformity, a narrow vocabulary band, heavy use of connectives like “moreover” and “furthermore”, and very few idiosyncratic choices.

Now think about how a strong ESL writer is often taught. Use full sentences. Avoid contractions. Link your paragraphs with explicit signposts. Prefer the standard word over the colloquial one. Every one of those habits pushes your text toward the machine-like end of the distribution — not because it is machine-made, but because the same advice that improves clarity also removes the irregularities detectors associate with human writing.

That is the crux of the false positive. The detector is not detecting you. It is detecting a style, and your style sits close to a style that models also produce.

There is a second layer. Detectors are trained and tuned on English text, usually edited English. When your writing carries transfer features from your first language — different article usage, a slightly different rhythm, a preference for nominalisation — some tools handle it well and some do not. Two detectors can read the same paragraph and disagree sharply. That disagreement is itself useful information, and it is worth understanding before you panic. Our own explainer on why two tools disagree walks through the reasons.

If you want the mechanics rather than the reassurance, how AI detection works covers the signal families in detail.

What our own measurement showed about flagged human writing

We run our own detector on our own writing, which is an uncomfortable exercise but a useful one. On 2026-09-19 we scored 46 hand-written DeAIze guides with our local ensemble at a 30% flag threshold. The mean AI score was 27%, the median 25%, and the range ran from 0 to 55%. At that threshold, 17 of the 46 were flagged.

Read that again. These are guides written by people, for people, on a site that sells a humanizer. More than a third of them crossed the line at a threshold that a cautious institution might use.

In the same run, 12 unedited passages generated from fixed prompts scored a mean of 35% and a median of 34%. The gap between human-written and machine-written text in that sample is real but narrow — nowhere near the clean separation people imagine.

The most interesting number was the AI-tell frequency: 3 per 100 words in the human sample versus 1 per 100 words in the unedited model output. That is not a typo. Our human writers used more of the phrases commonly labelled as AI tells than the model did. The tells are a folk taxonomy, not a reliable test, and teachers who mark essays by counting “delve” and “moreover” are working from a list that does not survive contact with data.

Treat all of this as what it is: our own instrument, on our own corpus, on one date. It is not a study of Turnitin, GPTZero or anyone else. But it does show why a flag deserves a conversation rather than a verdict.

If you want to see which specific sentences in your own draft carry the signal, run it through the sentence-level detector before you rewrite anything. Knowing the three lines that triggered the flag is far more useful than knowing the overall number.

Why do ESL writers get flagged more often?

Three mechanisms do most of the work, and they compound.

Training-data asymmetry. Detectors learn what “human” looks like from corpora that are overwhelmingly edited native-speaker English. Your writing may be perfectly good English and still sit outside that reference distribution.

Register flattening. Academic English taught to second-language learners is deliberately neutral. Neutral register is exactly what instruction-tuned models default to. You were taught to write like the model writes.

Uniformity penalties. Many detectors score low variance as a machine signal. A writer who has been drilled into consistent paragraph structure produces low variance on purpose.

None of these mechanisms involves intent. That is the point to hold onto when you are sitting in a meeting being asked whether you used ChatGPT.

How do you raise it with a teacher or an institution?

The instinct is to lead with “I didn’t cheat.” Resist it. Denial invites a debate about your character. Instead, ask about the process the institution uses and offer your own evidence. A short, calm message does more than a long, wounded one.

1. Ask what standard applies. Which tool, what threshold, and what weight the score carries in the policy. Many institutions have updated their academic integrity policy without briefing students on the specifics.
2. Ask what would count as evidence of authorship. This reframes the conversation from accusation to procedure, and it usually gets you a straight answer.
3. Offer your process record before you are asked. Drafts, timestamps, notes, the reference list you built by hand.
4. Ask for a meeting, not an email thread. Tone is easier to read in person or on a call.
5. Say what you want. A note in the file, a chance to walk through your sources, or an oral defence of the argument. Specific asks are easier to grant than vague appeals.

One sentence is worth rehearsing: “I understand the report raised a concern, and I’d like to show you how the piece was built.” It is not defensive, and it hands the reader a path forward.

The fear underneath all of this is real, and it deserves an honest answer rather than a soothing one. A flag can follow you — into a grade appeal, a supervision relationship, a visa file, a first job reference. Institutions move slowly and records persist. That is precisely why you should not wait for the process to reach you: document as you write, not after you are accused.

What evidence is worth keeping from your drafting process?

You cannot reconstruct a process after the fact, so build the trail while you work. Keep it boring and automatic.

Evidence	Why it helps	How to capture it
Version history	Shows the argument developing over time	Save dated drafts; use your editor’s version feature
Source notes	Links your claims to things you actually read	A single running notes file per project
Outlines and mind maps	Shows structure before prose	Keep the messy original
Reading annotations	Shows your thinking, not a model’s	Highlighted PDFs, margin notes
Supervisor or peer feedback	Third-party corroboration	Save emails and comments
Oral explanation	You can defend the argument live	Prepare a two-minute summary

A version history that shows a paragraph getting worse before it gets better is more persuasive than a polished final file. Human writing is lumpy. Keep the lumps.

How do you edit without flattening your own syntax?

The wrong move, and the one people make under pressure, is to rewrite until the detector is happy. That usually means shorter sentences, simpler words, more hedging — the exact opposite of what makes your writing sound like you.

A better approach: keep your structure, vary your rhythm, and add the specifics only you can supply.

Read the flagged sentences aloud. If every sentence lands at the same length, break one in half and let another run long. Rhythm is the hardest thing to fake and the easiest thing to restore.

Replace one generic phrase per paragraph with a concrete one. “The results were significant” becomes the actual measure and the actual direction. This is not padding; it is precision, and it is the single fastest way to sound like a person who was there.

Keep your own connectors, including the ones a style guide would cut. If you naturally write “and so” instead of “therefore”, leave it. Your quirks are evidence.

Do not swap words for synonyms just to move a score. Paraphrasing to dodge a detector usually produces worse prose and often does not move the number much. The difference between the two approaches is worth reading about in paraphrasing vs humanizing.

Here is an illustration, not a real student’s work. Before: “Furthermore, the implementation of renewable energy policies has been demonstrated to be effective in numerous contexts. Moreover, these policies contribute significantly to the reduction of emissions.” After: “Renewable energy policies work, and we have the receipts. Germany’s feed-in tariff is the case I keep coming back to, because it survived a change of government.” Same argument, same author, different amount of evidence that a person wrote it.

If you want a structured pass rather than ad-hoc editing, cleaning up machine-made patterns gives you a checklist that does not require you to change your voice.

The trap to avoid is treating a humanizer as a laundering step. It is not, and using it to pass off work you did not do is academic fraud regardless of what any tool reports. Where a humanizer genuinely helps is a different situation: you wrote the draft yourself, in your own second language, and a tool is misreading your style. That is the case where a draft editor that rewrites only the flagged sentences — and checks that your meaning survived the rewrite — is doing something defensible.

What can a detector score actually establish?

Very little on its own, and this is the part worth saying plainly to anyone who waves a report at you.

A detector score is a probability estimate produced by a statistical model comparing your text to patterns it has seen. It can tell you that a passage resembles model output on the features that model measures. It cannot tell you who typed the words, when, with what assistance, or with what intent. It has no access to your browser history, your drafts or your brain.

The practical consequences:

• A high score is a reason to look, not a conclusion.
• Two detectors on the same text can disagree, and neither is lying.
• A low score proves nothing either, because editing to lower a score is easy and authorship is not the same question.
• The threshold your institution uses is a policy choice, not a scientific constant. A different threshold would flag a different set of students.

Our own numbers make this concrete. In the 2026-09-19 run, human-written guides averaged a 27% AI score and unedited model output averaged 35%. If you set a threshold in the middle of that band, you will catch some machine text and you will catch a lot of people. Anyone who tells you detectors are close to definitive is not reading the same numbers we are.

If you are dealing with a specific report, the breakdown of what a score means in practice is covered in false positive AI detector: what to do.

Ready to see which sentences read machine-made?

Before your next submission, spend ten minutes doing what an appeal would force you to do anyway: find out which lines carry the signal, and decide whether they are worth changing.

Paste your draft into the AI detector to get a document score plus a per-sentence breakdown, so you are reacting to three specific lines rather than a single frightening number. If the flagged sentences are ones you genuinely wrote and want to keep, the draft editor rewrites only those lines, keeps a change only when the meaning survives, and reports the before and after. You can see the current pricing — pay-as-you-go, credits that do not expire, and a free tier of 200 words on signup with no card required. If you already know what you want to test, create an account and run your own paragraph first.

And whatever you use, keep the drafts. The strongest defence against a false positive is not a lower score. It is a folder that shows how the piece was built.

What does a sentence-level score actually point you to?

A single document number tells you a flag happened. It does not tell you where. That distinction matters when you are deciding whether to edit at all, because most flagged drafts contain a handful of lines carrying the signal and a lot of ordinary prose that never triggered anything.

Sentence-level detection splits the document into passages and scores each one, so you can see which lines read machine-made and which do not. When you look at that breakdown, three patterns tend to show up:

• A cluster in one section. Usually the most abstract paragraph — the one where you summarised rather than described. That is an editing problem, not an authorship problem.
• Scattered single lines. Often your transitions and topic sentences, the ones drilled into you as correct academic form.
• A high document score with no strong passage. This happens when many mid-range lines add up. Nothing looks damning on its own, which is worth knowing before you rewrite anything.

For a worked walkthrough of what each score definition means and where a flag sends you next, see DeAIze review: what the detector actually reports.

Our own run on 2026-10-13 scored 1549 passages from DeAIze guides not marked generated (metadata selection only; human authorship not independently verified) at a 30% threshold. The mean AI score was 28%, the median 20%, and 522 of those 1549 passages sat at or above the threshold. Passage counts, not documents, and not a verified false-positive rate. The point stands: a threshold catches a lot of ordinary writing, so read the per-sentence view before you decide what to change.