Because a detector doesn't measure who wrote something — it measures how predictable the writing is. Clean, formal, well-structured academic prose scores as machine-like, because that's the style language models were trained to imitate.
Which means the habits your marking rubric rewards are the same ones that raise a detector's score. You can be flagged for writing well.
The mechanism, in one paragraph
A detector runs your text through a language model and asks, at every word: how surprising is this, given everything before it? Low surprise across a whole document reads as machine-generated, because models pick likely words by design. That's the entire measurement. There's no watermark being read, no database being matched — just a running score of how expected your prose is.
Nothing in that process can distinguish “a model wrote this” from “a careful person wrote this in plain, correct, conventional English.”
What raises your score without you doing anything wrong
- Even sentence rhythm. You were told to be concise and consistent. Consistency is low surprise.
- Formal academic register. Hedged, impersonal, cautious phrasing is the most predictable prose there is.
- Clear signposting. Topic sentences and explicit structure — required by most rubrics, and highly patterned.
- Writing in a second language. Safe, correct, well-drilled constructions rather than risky idiom. This is the biggest single factor: a Stanford study found roughly 61% of essays by non-native English speakers were wrongly flagged. See this answer.
- Quotation and technical terminology. Quoted passages and fixed disciplinary phrasing are, by definition, text that already exists. More in can quotes trigger a detector?
- Editing tools. Accepting a lot of grammar suggestions pushes prose towards the conventional. Covered in does Grammarly make my essay look AI-generated?
Drafts written quickly and messily, then lightly tidied, often score lower than carefully polished work. The students most likely to be flagged are frequently the ones who worked hardest on the prose.
This is a known failure, not a fluke
HEPI spent the summer of 2026 publishing on it, concluding that a detector score should never be sufficient evidence of misconduct on its own. The Office of the Independent Adjudicator upheld complaints that year from students — including non-native English speakers and an autistic student — accused on a score and later cleared. Some universities abroad have switched their detectors off. If your essay was flagged and you wrote it, you are in a large and well-documented group.
What to do about it
If you've been accused, the process matters more than the argument: start here. If you haven't and you're worried, the useful move isn't to write more erratically to fool a tool — it's to keep your drafts and version history in one continuously-edited document, so the question can be answered with evidence rather than opinion.
SafeGrade points at the specific sentences raising flags so you can rewrite them in your own voice. Analysed privately: never stored, never used for training.
Run one free scan →No card. No catch. First scan free.