How we detect Claude text
Our classifier is trained on human writing and on text from about 30 current AI models, including models from Anthropic. It combines two independent signals: how predictable each word is to a language model, and the style of the passage. It does not depend on spotting a few telltale phrases, so it keeps working as Claude changes.
See which sentences read like Claude, not just one overall score.
Thresholds are calibrated on human English and Chinese writing.
Export the result as a PDF to keep or share (paid plans).
Check text from your own app with the JSON API.