humanflow

Sentence-level detection

Sentence-level detection reports a score for each sentence rather than one figure for the whole document — the same measurement at a resolution that shows which passages produced the result.

Last reviewed 15 August 2026 · The HumanFlow team

In plain English

Instead of "this essay scores 40%", you get to see which sentences the tool objected to.

That does not make the measurement more accurate. It makes it inspectable, which is a different and often more useful property.

A worked example

The same 40% essay from the previous entry, reported per sentence. Only the resolution has changed.

  ¶1  Social media has changed how people communicate.        ▓▓▓▓▓  high
      These changes have been documented by researchers.      ▓▓▓▓▓  high
  ¶2  Teenagers in the 2021 cohort reported both, often in    ░░░░░  low
      adjacent questions on the same form.                    ░░░░░  low
  ¶3  [block quote, cited]                                    ▓▓▓▓░  high

  document score: 40%  ← the same number as before

The number did not change; the situation became legible. Two opening sentences of generic scene-setting and one correctly cited block quote produced almost all of it.

That is a materially different conversation from "forty percent of this essay". The flagged material is a weak introduction and a quotation, neither of which is misconduct, and both of which are visible in one glance.

It also gives a writer something to act on. The high-scoring sentences are usually the ones worth rewriting on their own merits, because flat generic prose reads as flat generic prose to a person too.

Why it matters for AI detection

Because it converts a verdict into evidence. A per-sentence view can be examined, disagreed with, and checked against the draft history — a document-level percentage can only be believed or not.

It is also the honest way to present an uncertain measurement. Showing where a model's confidence sits, rather than averaging it into one figure, makes the tool's limits visible rather than hiding them behind a decimal point.

The caution is that finer resolution is not better accuracy. A sentence-level view built on the same unreliable signal is the same unreliable signal, shown in more places. It tells you where the score came from, not whether the score is right.

Commonly confused with

Document-level detection
The same measurement reported at different resolutions. Neither is more accurate; one is inspectable and the other is not.
Stylometry
Sentence-level detection scores passages against a model's expectations. Stylometry compares a whole text against a known author's habits and needs a reference sample.

Read next

Part of the AI detection glossary.