Detection accuracy

How well our own detection model tells human and AI writing apart, measured on documents it never saw during training. We publish these figures for every model we put into use.

How to read these figures

• False positives first. We set each language's threshold so that at most 1 in 100 human documents in the validation set is flagged, then report how much AI text is still caught at that point.
• Test data is not your data. Human texts come from Wikipedia and web text written before 2021; AI texts were written by current models, including AI-polished and "humanized" texts. Formal, translated, non-native or very short writing can behave differently.
• One signal among several. The final result also uses language-model readings and forensic evidence, and never says "AI-generated" from one weak signal.

A certificate records the results of an automated analysis at a point in time. AI-detection and plagiarism results are probabilities, not proof, and a certificate does not certify authorship.