The mechanism, in one paragraph
Detectors score two things. Perplexity asks how surprising each word is given what came before. Burstiness asks how much sentence length and complexity vary across a passage. Language models pick likely words and produce even rhythm, so their output is low on both. A detector that sees low perplexity and low burstiness reports "AI" — regardless of who actually typed it.
Who this happens to
Not randomly distributed. The people most likely to be flagged for work they wrote themselves are, in rough order of documented risk:
- Non-native English speakers. The largest and best-documented bias. A smaller active vocabulary and simpler sentence construction produce a low-perplexity signature. Stanford researchers found seven detectors averaged a 61.22% false positive rate on TOEFL essays — and near-perfect accuracy on essays by US 8th-graders.
- Students writing to a taught structure. The five-paragraph essay is close to an algorithm. Following it well produces exactly the uniformity detectors penalise.
- Technical, legal and scientific writers, where conventional phrasing is required by the field and originality of expression is actively discouraged.
- Careful editors. This one is genuinely perverse: grammar checkers, style tools and diligent self-editing all reduce variance. Polishing your prose makes it look more machine-written.
- Autistic and neurodivergent writers, who may write in more consistent, structurally regular patterns. Less formally studied than the ESL bias, but reported widely enough to take seriously.
A worked example
These two paragraphs say the same thing. The first is the kind of prose that gets flagged; the second usually is not. Neither was written by a model.
The industrial revolution was a significant period of change. It transformed manufacturing processes and had a major impact on society. Many people moved from rural areas to cities during this time. This migration changed the structure of communities in important ways.
The industrial revolution gutted the village. Between 1750 and 1850 whole families walked off land their grandparents had farmed and into Manchester's cotton mills — and the communities they left behind never really recovered, because the people who leave first are the young ones.
The first is uniform: four sentences of similar length, abstract nouns, no specifics. The second varies its rhythm, names a place and a date range, and carries an argument. That difference is most of what a detector is measuring. It is also, not coincidentally, the difference between writing that is dull and writing that is good.
If your work has been flagged
- Gather your draft history first. Google Docs keeps full version history under File → Version history. Word keeps AutoSave revisions. This is the single most effective response available, because it shows the work forming over time and a score cannot rebut it.
- Ask which tool was used and what the threshold was. Vendors publish guidance saying scores should not be the sole basis for an accusation. Turnitin's own documentation says this.
- Run the same text through two or three other detectors. They routinely disagree. Disagreement between tools on identical text is itself an argument about reliability.
- Point to the research. If English is not your first language, the Liang study is directly relevant and was published in a peer-reviewed journal.
- Offer to discuss the content. Someone who wrote a piece can talk about the choices in it. That conversation persuades a human reviewer more than any counter-score.
What we can and cannot tell you
We can tell you that false positives are real, common, and unevenly distributed. We can point you at the research. What nobody can honestly tell you — including any tool promising it — is that a given piece of text will score a given way with a given detector. Detectors update, they disagree with each other, and none of them publish reproducible accuracy figures.
The full accuracy picture, with the numbers and their limits.