Why Your Writing Reads as AI — and Why That Is Not Proof

Every AI detector answers the wrong question. «This text is 87% AI» is a claim about a probability distribution, not about you, and it cannot be checked, appealed or disproved. Worse, the number is systematically wrong for people who learned English as a second language: the Stanford study behind most of the 2026 coverage found that over 61 per cent of essays written by non-native English speakers were misclassified as AI-generated, against roughly 5 per cent for the native-speaker essays in the same study.

So this page does the opposite. It shows you the specific sentences that carry the signals detectors weight most heavily, and names the habit behind each one. You end up able to say «this paragraph reads flat because I was taught to avoid contractions in academic English» — a sentence that survives a conversation with a professor. «The tool said 12%» is not.

Detectors do not recognise machine authorship. They measure how predictable text is, and predictable is not the same as fake. Vendors do not publish their feature sets, so the list below is what the public research and the tools' own descriptions point at rather than a specification — but a careful second-language writer produces all four honestly:

• Even sentence length. Native prose lurches — a nine-word sentence next to a thirty-word one. Writers working in a second language tend to stay inside a safe range they can control, and that evenness is one of the properties detectors are known to weigh.
• No contractions. «Do not» rather than «don't». Most English-as-a-foreign-language instruction treats contractions as informal and drills them out; so does most model output.
• Connector scaffolding. «Moreover», «furthermore», «in conclusion». Academic English courses teach these as the skeleton of an argument. They also turn up disproportionately often in generated text.
• Repeated openers. Six sentences that begin «The» or «It is». Comfortable, correct, and flat.

Notice what is not on that list: errors. Detectors are not looking for mistakes. They are looking for smoothness, which is exactly what a diligent non-native writer works hardest to produce.

Two things are worth knowing about where this is heading. As of August 2026 at least a dozen universities have switched Turnitin's AI detector off entirely, and others — Yale, Vanderbilt, Johns Hopkins, Indiana among them — now forbid treating a detector score as the sole evidence of misconduct. Find out which of the two positions your institution holds before you argue, because only the first makes the score inadmissible.

But the replacement is not neutral either. Process-based assessment — staged drafts, oral defences, writing-process analytics that log every paste — reads a perfectly legal bilingual workflow (draft in your first language, translate, refine) as a suspicious edit history. Against a process record, «the score is unreliable» is no longer an answer, because a process record is not making a statistical claim. The only counter is a better record of your own, made while you write rather than reconstructed afterwards.