Ask someone how much of a document is AI-generated and they will usually answer with a percentage, as though text were blended like paint. In practice, documents are assembled.
A typical hybrid document starts as a human outline, gets expanded by a model, is edited by a person who disagrees with half of it, and ends with a conclusion the person wrote themselves at 11 p.m. Some paragraphs are wholly human. Some are nearly untouched model output.
Single-score detection handles this badly, and it is worth understanding why.
Consider a 1,200-word report where 300 words came from a model. An honest answer is that a quarter of the document is generated. What a single overall score tends to produce instead is a number somewhere in the middle: 40%, 55%, 60%. That number is not wrong so much as useless. It describes a blend that does not exist anywhere on the page.
This is the situation that generates most of the arguments about detection accuracy. The tool answered the question it was asked. It was asked the wrong question.
Averaging hides the thing people actually need to know, which is where the generated text is.
Two documents can produce identical scores and have nothing in common. In one, every paragraph is slightly machine-flavored. In the other, nine paragraphs are clearly human and one was pasted in. The first is a writer who leans on a tool throughout. The second is a writer who got stuck on one section. The responses these two cases deserve are completely different.
When we look at hybrid documents, the same structures keep appearing.
Spotting which pattern you are looking at tells you what kind of help the writer actually needs.
Breaking the text into paragraphs and scoring each one independently turns a mystery number into a map.
You can see that the introduction scored 12% and the middle section scored 88%. You can see that the flagged paragraph is the one that reads like a general summary and contains no sources. You can point at the exact lines in a conversation with a student, a client, or a colleague, instead of arguing about an aggregate score.
The paragraph that scores low matters as much as the one that scores high. A human-written conclusion surrounded by flagged body text is strong evidence that a person did the thinking and the tool did the typing.
If you are checking your own work before submitting or publishing, use the breakdown rather than the top-line score.
If you are reviewing someone else's work, the same list applies, plus one rule: never quote a single score as evidence. Quote the paragraph and the reasoning.
Text is rarely uniform, and people deserve an assessment that reflects that. A paragraph-level view is slower to read and harder to weaponize, which is precisely why it is more honest than a percentage.