AI Content Detector

Every AI detector on the internet gives you a percentage and asks you to trust it. That is a strange thing to accept for a measurement with a known false-positive rate, and it is the reason this page shows the eleven numbers underneath the headline instead of only the headline.

A detector cannot prove provenance. What it can do is measure style, report which features of that style are unusual, and let you decide what to make of it. Whether the text in front of you is unusual because a model wrote it or because you write in a formal register is a judgment only you are in a position to make — but you cannot make it at all from a single opaque number.

Two kinds of evidence

The score has two independent halves, and keeping them apart is the whole design.

Wording is what most detectors measure: how often the words models reach for actually appear, how many stock phrases turn up, how dense the em dashes are, how many sentences open with Moreover, how much passive voice there is. It is easy to measure and easy to defeat — swap thirty words and the number drops.

Structure is harder to measure and harder to defeat. It asks whether the sentence lengths vary the way a person's do, whether consecutive sentences open the same way, whether the paragraphs are all one tidy size, and — the part that matters most — whether anybody is actually in the text.

The two are combined as "not both of them failed to show it", so the total never falls below either input. On the engine's own calibration set that ordering comes out right: hand-written prose scores 7, a rule-engine rewrite of model prose scores 29, raw model prose scores 85, and the same model prose with every flagged word scrubbed out scores 90 — higher than before, because stripping the vocabulary also stripped the human texture and left the structure carrying the verdict on its own.

That single comparison is why a synonym-swap humaniser does not work, and why this page bothers to show you both halves.

The eleven measurements

Each is displayed against the most it could have scored on your particular text rather than against a fixed maximum. That normalisation matters: on a 40-word blurb, paragraph shape, cadence, openers and texture all stay silent for lack of material, and reading the vocabulary hits against a ceiling that assumes they fired would let a short machine-written text look innocent for no better reason than that there was not enough of it to judge.

The baselines it measures against

None of these thresholds is invented here. They are what ordinary human prose does:

The last two are the interesting ones. People do not write sentences of even length, and the variation is not a stylistic choice so much as a symptom of thinking — a clause gets away from you, then you write four words to land the point. A coefficient of variation below about 0.42 means the lengths are being set by something other than the argument.

What the score cannot see

Read the bands as descriptions of style, not of authorship. Above 75 reads as almost certainly machine-written, 50 to 75 as a strong machine fingerprint, 30 to 50 as noticeable machine patterns, 15 to 30 as light traces, and below 15 as human.

Three groups get false positives constantly, and none of them is doing anything wrong. Writers working in a second language, who reach for taught connectives. Anyone writing legal, regulatory or academic prose, which is uniform and passive on purpose. And anybody scoring a fragment rather than a document — under 150 words the rates mean almost nothing, which is why the page labels short text as unreliable instead of quietly returning a confident figure.

The part only you can fix

Two of the structural terms are content rather than style: whether anybody is in the text, and whether anything in it could be pointed at. A first person, an opinion, a real number, a specific example, a question aimed at the reader.

No tool can add those honestly, because inventing an opinion or a figure is exactly what makes AI content bad. What the page can do is tell you how many points they are worth on your draft and list the sentences where adding them would land — which turns the least actionable part of the score into a two-minute editing task.

Scoring runs entirely in this tab, on up to 400,000 characters, with no account and no request to anything. Once you know what to change, the AI Text Humanizer makes the wording changes and scores the result beside the original, and the Readability Analyzer handles the half of this that is about difficulty rather than provenance.