AI detection false positives, by type of writing
Seven kinds of genuine human writing that score high, why each one does, and what it takes to move it.
Wrong flags concentrate rather than scatter. If you can recognise which of these describes your document, you know why the number came out the way it did and what would change it.
Turnitin's published wrong-flag rate on human writing is 4%. Ours is 0.4%, which is 99.6% specificity, measured on 15,900 documents of real academic writing.
1. Methods and procedure sections
Formulaic by design. Narrow vocabulary, conventional constructions, phrasing frequently reused from an earlier paper or a lab protocol. Everything a classifier reads as regular.
This is the single highest-scoring section type in academic writing and the score says nothing except that the genre works the way it works. See AI detection in academic papers.
2. Anything heavily edited
Each revision pass smooths the prose. Smoothing removes variation, and variation is the strongest human signal there is.
The uncomfortable consequence: a fifth draft scores higher than a first, and a piece that went through a writing centre scores higher than the one that did not. See does Grammarly get flagged as AI.
3. Writing by non-native English speakers
Repeatedly measured at elevated rates in published evaluations. Learned English applies its rules more consistently and reaches for the common construction more often, and consistency is exactly what these tools read as machine-like. See why non-native English writing gets flagged.
4. Structured abstracts and short assignments
Short and templated at the same time, which is the worst pairing available. Every detector is substantially less reliable under a few hundred words, because sentence-length variation needs a sequence to exist at all. See why detectors are unreliable on short text.
5. Literature reviews
Long, densely hedged, structurally repetitive by necessity, and usually the most revised part of a document. All four properties push the same way.
6. Technical and legal prose
Field conventions constrain word choice hard. When the correct term is the only term, word choice becomes predictable, and predictability is what perplexity-based signals measure. See perplexity and burstiness, explained.
7. Human writing sitting next to assisted writing
Turnitin found 54% of falsely flagged sentences sit directly beside a sentence scored as AI written, and 26% two sentences away.
Four in five wrong flags land next to a genuine one, so a document with a single assisted paragraph shows a smear across the human writing around it rather than one clean highlight. See documents that are part human, part AI.
What moves any of them
The same short list, none of which touches what you are saying:
Vary sentence length deliberately. Our corpus puts human writing at a standard deviation of 8.7 words against 6.4 for generated text. Split one long sentence, merge two short ones, and the number moves.
Cut connective openings. "Moreover", "Furthermore", "Additionally" open 2.4% of human sentences and 5.8% of generated ones.
Put the specific detail back. Name the method, give the number, use the actual example. Generalising is the habit that reads as generated.
Full list in what makes text read as machine written.
Ours is free and unlimited with no account, scores every paragraph separately so you can see which of the seven applies to which section, and never stores your text.