Document-level detection
Document-level detection reports a single score for an entire document rather than for its individual parts — the format most tools default to, and the one that tells a reader least about why the number is what it is.
Last reviewed 15 August 2026 · The HumanFlow team
In plain English
One file in, one percentage out. Whatever produced that percentage is averaged away before you see it.
It is the cheapest thing to report and the hardest thing to act on, because it points at a document rather than at a passage.
A worked example
A five-paragraph essay scoring 40%, and the four readings of that number that are all still possible afterwards.
document score: 40% could mean two of five paragraphs are heavily flagged could mean all five are moderately flagged could mean the methods section alone is flagged could mean the quoted material is flagged the score is identical in all four cases
Those are four different situations requiring four different responses, and the reported number cannot distinguish them.
The fourth case is the one that causes the most avoidable harm: block quotations and reference lists are formulaic by construction, so a document can score on material the student did not write and correctly attributed.
Nothing about the score is wrong here. It is doing what it was asked to do, at a resolution that discards the information needed to interpret it.
Why it matters for AI detection
Because a number you cannot locate is a number you cannot check, and appeals turn on locating things. "Forty percent of this essay" is not an allegation anyone can respond to; "these two paragraphs" is.
It also concentrates the false-positive problem. A document containing one formulaic section and four pages of distinctive writing can be reported at a level that reads as serious, because the average does not care which part earned it.
If you are given a document-level score and nothing else, asking which passages produced it is a reasonable first question, and often the fastest route to resolving the matter.
Commonly confused with
- Sentence-level detection
- Same underlying measurement, different resolution. One reports a figure for the file; the other reports it per sentence, so you can see which passages drove it.
- AI writing score
- The AI writing score is usually the document-level number. This term names the reporting choice that produced it.
Read next
Part of the AI detection glossary.