This checker looks for worn, formulaic phrasing that has become strongly associated with machine-drafted writing: words like delve, patterns like "It's not just X, it's Y", throat-clearing openers, importance puffery, signposting, fake-profound endings and 20 more. It scores what the words are, not who wrote them.
It does not and cannot tell you whether a text was written by a person or a model. Some human writing is sloppy; some machine writing is clean. Anyone using this output to accuse a person of using AI is misusing the tool.
The same text always gets the same score. No model call, no randomness, no stored text. The engine is a port of the same rule set used to train and gate our Clotho 7B model family: every training pair we accepted had to pass the chosen side of these gates, and every rejected side had to fail them.
Each finding has a severity: high (always slop: banned words, weasel attribution, importance puffery), medium (formulaic patterns), low (style choices). The weighted score is high x 3 + medium x 2 + low x 1. Verdicts: clean = no high findings and weighted under 10; needs work = no high findings but weighted 10 or more; slop = any high finding fires.
Free tier: 2,000 characters per check, 10 checks per minute per IP. The engine is tuned on English prose; other languages will under-report. Rules are a floor, not a ceiling: clean text is not automatically good, only free of the patterns we test for.