An AI content detector measures style. Provenance is a different question.

Every tool sold in this category answers a question it does not have the evidence to answer: who wrote this. It has no writing session, no keystrokes, no model logs, no author's notes. It has the finished text and a training set labelled human and machine, and from that it produces a number.

The number is a description of style. Style is real, measurable and worth knowing about, but it belongs to the document rather than to a person. A document can be stiff, generic and rhythmically flat whether a model wrote it, a committee wrote it, or a careful writer produced a first draft at midnight and never edited it.

Keeping measurement and accusation apart is the discipline. A style score tells you how the prose behaves. Authorship is a claim about a human being, and a claim about a human being needs evidence that survives someone disagreeing with it, which a percentage does not.

Supervised classifiers are the most common. A model is trained on examples labelled human and AI, then asked to place new text on that boundary. They are cheap to run, sensitive to the label quality of their training data, and quick to drift when the generators change.

Statistical heuristics measure properties of the prose directly: token predictability, sentence-length variation, vocabulary range, punctuation habits, pronoun density. They are explainable, which is their advantage, and they confuse style with origin, which is their limit.

Provenance methods are the only family that genuinely addresses origin. Metadata, content credentials and vendor watermarking do not guess at authorship, they carry a record of it. Coverage is thin, because most publishing pipelines strip metadata and most models do not watermark reliably, but when provenance data exists it is worth more than any score.

Different training sets, different thresholds, different tokenisers, different handling of length. Add the commercial decision about where to place the cut-off and you have two tools that will disagree on the same paragraph, both with a confident percentage beside them.

Disagreement is not a scandal. It is what happens when a soft measurement is presented as a hard one. Two thermometers in the same room read within a fraction of a degree. Two authorship detectors can differ by fifty points, which tells you something about the precision of the underlying signal.

Treat any single score as a rough editorial signal and the disagreement becomes informative rather than confusing. When one tool flags a draft hard and another does not, read the draft. The paragraph that triggered the first one is usually the paragraph worth editing anyway.

Detectors are more confident on text with predictable syntax. That describes legal boilerplate, safety instructions, compliance updates, press releases, product specifications and plain-language versions written for accessibility. Those genres score high because they are formulaic, not because a machine wrote them.

It also describes a large amount of work from writers writing in a second or third language, whose academic training encouraged longer sentences and formal connectives. The same properties that make such text clear for an international audience make it look machine-made to a classifier.

Students, translators, non-native colleagues and anyone writing in a house style have all been on the receiving end of a confident accusation built on a number nobody could audit. If your process can produce that outcome, the process is the problem, not the tool.

Triage. A hundred drafts a week cannot all be read closely, and a style measurement is a reasonable way to sort them so the weakest get attention first. No accusation, just an order of work.

Editing feedback. The measurements behind a detector are the same ones an editor would make with a pencil: repetitive openers, uniform sentence length, abstract nouns where a name belongs, claims with no evidence. A detector that exposed its sub-scores would be a useful editing instrument even if it never mentioned authorship.

Consistency at scale. When a team publishes across twenty pages, a style check reveals when one section was assembled differently from its neighbours, which is a maintenance problem worth catching early.

Five properties do most of the work, and all five are editable.

Detectors need volume. Below roughly two hundred words, variance is dominated by noise and the score is close to meaningless. A product blurb, a caption or a two-line reply cannot be classified with any reliability, whatever the interface implies.

Translated copy is its own trap. Translation flattens idiom, standardises syntax and produces the same even cadence across paragraphs, so a translated page often scores higher than the original it came from.

Heavy editing moves text the other way. Once a writer has cut the scaffolding and varied the rhythm, a detector loses its grip, which is the strongest argument that the tool is measuring style. Improve the prose and the number changes, without anyone's authorship changing at all.

Write down what the check measures before you run it, and what you will not do with the result. Then keep the same rule for employees and vendors, because a rule applied only to contractors is a different rule.

Record the version of the tool, the text that was submitted, the score and the date. If the decision is ever questioned, an auditable record is worth more than a screenshot of a gauge.

Never publish a percentage as proof. A sentence that says the prose showed a high density of stock phrasing is defensible. A sentence that says the writer used AI is an allegation, and the tool cannot support it.

Route disputed cases to a named human who reads the work and owns the answer. That is not a courtesy, it is the only step in the process that can carry the weight of a judgement.

Does this sentence give a reader something they can act on? A name, a number, a consequence, a decision. Prose built from specifics seldom scores high, because specifics are not drawn from an average of everything ever written.

That question is also more useful than detection to the person who paid for the work. They do not want to know whether a model helped draft the page. They want to know whether the page will earn trust with the audience it reaches.

So run the style check, read the flagged paragraphs, edit them, and keep a record. Detection stops being a lie detector and becomes what it should have been: a lint pass for prose.