An AI phrasing detector measures prose, not provenance.

Detectors that claim to identify AI text are answering a provenance question with a style measurement, and the mismatch is why they disagree with each other and with reality.

Phrasing analysis avoids that error by not making the claim. It reports specific, countable properties of the prose, and lets an editor decide what they mean for this publication and this audience.

That distinction matters legally as well. A style score is a reasonable editorial input; it is a poor basis for accusing a writer.

Five measurements hold up across formats, and each points at a different kind of problem.

Readers do not experience word choice the way editors do. They experience rhythm, and rhythm is variance: a short sentence landing after a long one, a fragment used deliberately.

Uniform length is the strongest visible signature of generated prose, because a model produces sentences of similar structural weight. Averages hide it, which is why the check reports spread rather than mean.

The fix is editorial and quick: cut one sentence to four words, let one run long, delete the transition that announces what the next paragraph will do. Variance returns and the paragraph starts sounding like a person.

Legal writing should hedge, because qualified claims are accurate claims. Technical documentation should use the same noun for the same thing, which looks like repetition. Safety copy should be absolute where safety is absolute.

So the numbers are relative to the format. A report that treats every hedge as a defect will make legal copy worse, and a report that treats every absolute as an overclaim will make an instruction manual unreadable.

Set expectations per format rather than globally. Marketing landing pages have the highest tolerance for consequence-free prose; regulated claims have the lowest.

Read the flagged paragraph and ask what it asserts. If the answer is nothing, delete it, and the counts fall without any word-level surgery.

If it does assert something, replace the stock verb with the specific action and the abstract noun with the thing itself. A sentence with a name, a number or a consequence in it is rarely flagged, because specifics do not come from a distribution.

Then re-run. The comparison between the first and second run is the useful artefact: it shows whether the editing changed the prose or merely rearranged it.

No measurement of phrasing can tell you whether the argument is honest, whether the joke lands or whether a claim will be read as a promise by the audience it reaches.

For anything published under a company name, in a regulated sector or in front of a legal team, the numbers narrow the reading and a named editor signs the verdict.

That is the division of labour the tool is built for: measurement on everything, judgement where judgement is needed.