Similarity Is 2% but the AI Score Is 21% — How to Read That Combination

A 2% similarity score and a 21% AI score are not in contradiction, because they measure different things. Similarity asks: does your text match other published text? The AI score asks: does your text read statistically like machine-written prose?

Both answers can be true at once: your text matches almost nothing (that is why similarity is 2%), and your text reads like machine prose in some passages (that is why the AI score shows 21%). The low similarity does not clear you of the AI score — the two systems are independent.

And 21% is significant, but not for the reason people think. It is the first number that actually appears on the report instead of an asterisk. That is a display rule, not a verdict.

The two reports answer two different questions, and the combination you received is the standard shape of a student who wrote their own work but in a style the model reads as machine-like.

The similarity report compares your submitted text against Turnitin's collection of published content. A 2% score means your text matches almost nothing in that collection — not a sign of copying, and not a sign of AI either.

The AI writing report estimates how much of your text reads like machine-written prose, based on statistical properties of the writing. A 21% score means the model judged roughly a fifth of your qualifying text as likely AI-written.

Two separate questions, two separate answers. This is not a bug and not a contradiction.

The tempting interpretation is: "my similarity is 2%, so I clearly didn't copy anything — so why is the AI score high?" The answer is that the two scores do not speak to each other.

Turnitin's own documentation describes the AI score as independent of the similarity score. The similarity system looks for text that matches existing published content. The AI system looks at how your words are chosen — the statistical shape of your prose. A text can be 100% original (nothing matches anywhere) and still be statistically similar to machine writing. Originality in the copying sense and AI-likeness in the statistical sense are different axes.

So a low similarity score is not a certificate of "not AI". It is a certificate of "not copied". Both are useful, and neither replaces the other.

21% is the first number that actually appears. The reason matters more than the number itself.

Turnitin's display rule: to avoid potential incidence of false positives, no score or highlights are attributed for AI detection scores in the 1% to 19% range. When AI is detected below the 20% threshold in the report, it is indicated with an asterisk (*%) and no percentage is attributed. Above that display line, the report shows a number and highlights.

So 21% is not a verdict, it is the first visible number. A 14% score would be invisible — shown as an asterisk — and a 21% score is shown as "21%". The jump from asterisk to number is a display rule, not a doubling of guilt.

The official false-positive framing is worth knowing alongside it: Turnitin says it strives to keep the false positive rate under 1% for documents with over 20% of AI writing — in other words, it might flag a human-written document as AI-written for one out of every 100 fully human-written documents. Even in the band where the number shows, the official claim is that about one in a hundred fully-human documents could be flagged. A 21% score is in that band, so it is a signal to look at, not a sentence.

The 21% is not distributed evenly across your document. It comes from specific passages.

Turnitin describes the mechanism: sentences from the submission are extracted and segmented into overlapping sections for prediction analysis. Each segment gets a probability between 0 and 1; each qualifying sentence inherits the score of the segment it sits in. Because segments overlap, some sentences carry multiple scores, which are pooled into the document percentage.

Two practical consequences. First, the passages that carry the score are identifiable — if your instructor shares the report, the highlights show exactly which passages the model judged likely AI. Second, editing effort should go to those passages, not scattered across the document.

One caveat for short documents: if your submission is only a few hundred words, the prediction is mostly "all or nothing" — the model predicts on a single segment without overlap, and text that is a mix of AI-generated and original content could be flagged as entirely AI-generated. For a short paper, the 21% deserves a heavier discount than for a long one.

The conversation changes when you can point at the official wording.

Turnitin's own documentation says the percentage on the AI writing indicator should not be used as the sole basis for action or a definitive grading measure by instructors. That is a sentence you can quote, and it is not a loophole — it is the vendor's own position.

What to bring to the conversation: the number (21%), what it means (first visible number above the asterisk band, not a verdict), the location (the highlighted passages, if you can get the report PDF), and the mechanism (the score comes from specific segments, not from the whole document). The combination you are explaining is: my text is 2% similar to anything published, and 21% AI-likeness on a statistical axis — both can be true for an original, carefully written paper that uses formal academic phrasing.

Ask for the report before you argue. Students cannot see the AI indicator themselves — only instructors and administrators can — but instructors can download and share it as a PDF. The report turns a vague accusation into a list of passages you can address.

A 2% similarity score and a 21% AI score are not contradictions: one measures copying, the other measures statistical AI-likeness. The low similarity does not clear you, and the 21% is not a verdict — it is the first number that appears above the asterisk band, in a range where the official false-positive claim is under 1%. The score comes from specific passages, and the official guidance says it should not be used as the sole basis for action. That is the combination to bring to your instructor: two independent numbers, one explainable mechanism, and the vendor's own words on your side.