Detection
Writing-pattern analysis (stylometry)
The on-device text check scores sentence rhythm, overused vocabulary, lists of three, em dashes, markdown in comments and more. Every triggered pattern is listed in the card.
Updated
Text has no watermark anyone outside the generator can read, so writing is scored on patterns. The analysis runs inside your browser before anything is sent anywhere, and paints the circle immediately.
The features
Each feature returns a reading and a weight; the score is a clamped weighted sum, and every feature that triggers becomes a note in the card's Text pattern row.
- Burstiness. Human writing varies sentence length; models tend toward even sentences.
- Vocabulary models overuse. "Delve", "tapestry", "testament", "it's important to note", "in conclusion" and similar.
- Lists of three and parallel structures.
- Em-dash and curly-quote density.
- Markdown headings or bold labels inside comments, and emoji-bullet lists.
- The hedge-then-summary shape.
- A suspicious absence of typos and slang.
- Sign-off phrases.
- Type-token ratio, a measure of vocabulary variety.
Text under 40 characters is scored 0.1 and noted as too short.
How it combines
With no provenance, classifier or recheck signal, the pattern score is the verdict. Once three or more community reports exist it is blended with them: 75% pattern, 25% report ratio. The verdict is ai at 0.7 and uncertain from 0.4. See How verdicts are scored.
Treat it as a best guess
Careful human writers trip these patterns, especially in formal or edited text, and a model prompted to write loosely evades them. A text flag is evidence, not proof, and the terms prohibit using a verdict to accuse a specific person. If a flag is wrong, choose Not AI.