AI Writing Pattern Checker — Stylometric Analysis

No. Nothing can, and the tools that claim a percentage are overstating what stylometry supports. Every signal measured here is also produced by careful editing, by formal registers, by non-native English writers and by anyone taught the same essay structure a model was trained on. Published false-positive rates are high enough that acting on a score alone — in grading, hiring or moderation — is indefensible.

The variation in sentence length across a piece of writing. Human prose alternates between long and short sentences much more than generated prose does — a coefficient of variation near 0.55 is typical for people, and near 0.30 for models. It is the single most durable of these signals because it survives light editing.

Because a plain textarea produces a hyphen and a generator produces an em dash. It is deliberately weighted very low, since word processors, publishing tools and anyone with a compose key produce them too. It is a hint, not evidence.

This one is pure arithmetic over the text and runs instantly. The on-device detector loads a real language model and measures how predictable your text actually is to it — the perplexity method the published research uses. They measure different things and are worth reading together.