The similarity checker is easy to understand: it compares your text against a database and shows you the matches. The AI detector is a different animal entirely. There is no database of AI writing to match against — your essay is brand new, and so is every essay ChatGPT produces. So what is Turnitin actually measuring when it says 34% of your paper is AI-written?
It's a prediction, not a lookup
Turnitin's AI detector is a classifier: a model trained on large amounts of human writing and AI writing, whose job is to say which one a given passage looks more like. It breaks your document into overlapping segments of prose, scores each one, and aggregates the results into the document-level percentage an instructor sees.
What the model keys on — in general terms, since Turnitin doesn't publish its internals — are the statistical fingerprints of language-model output. AI text tends to be reliably predictable: each next word is a high-probability choice, sentence lengths cluster in a narrow band, and the rhythm stays even from paragraph to paragraph. Human writing is messier — we make odd word choices, write a seven-word sentence after a forty-word one, and drift in tone. Detectors measure that predictability and evenness, not meaning.
What the score counts — and quietly ignores
The percentage isn't "share of the file that's AI." It's the share of qualifying prose the model flagged. Bullet lists, tables, code, references, and very short or fragmented text don't qualify and are excluded from the calculation. Turnitin also needs a minimum amount of prose (on the order of a few hundred words) to score a document at all — which is why very short submissions sometimes come back with no AI score.
One more quirk worth knowing: low scores display as an asterisk. When the detector's result lands under roughly 20%, Turnitin considers it too uncertain to state a number and shows *% instead. If you've seen a masked score on a report, that's what it means — we cover the whole report layout in how to read the AI writing report.
Why human writing gets flagged
Everything above explains the detector's known failure mode: it flags human writing that happens to be statistically "even." Independent testing has repeatedly found false positives — Turnitin itself claims a low document-level false-positive rate, but low isn't zero, and across millions of submissions the absolute number of wrongly flagged students is real. Writing styles that get caught in the blast radius:
- Formulaic academic prose. Five-paragraph structure, uniform topic sentences, "furthermore/moreover" transitions — the house style of a nervous undergraduate is uncomfortably close to the house style of a language model.
- Non-native English writing. Writers working in a second language often use safer, more predictable constructions — research (including the widely cited Stanford study "GPT detectors are biased against non-native English writers") has shown detectors flag non-native writing at higher rates.
- Heavily tool-polished text. Aggressive grammar-tool rewriting smooths out exactly the irregularities that read as human.
What this means practically
Three takeaways. First, the detector reads style, not honesty — so an innocent paper can flag, and you should know your number before your marker does rather than discover it in an accusation (if that ship has sailed, here's what to do when your original work gets flagged). Second, only Turnitin's own model predicts Turnitin's output — free checkers run different models with different opinions. Third, if you do pre-check and see flagged passages, the fix is revision with texture — vary sentence length, add concrete specifics only you would know — not mechanical rewording.
The detector will keep evolving; Turnitin updates it as new models ship. What stays constant is the asymmetry: your instructor sees the score and you don't. Closing that gap is the whole reason running the real check first exists.