The article you're commenting on actually states in its abstract: > Over a wide range of document types, Binoculars detects over 90% of generated samples from

The article you're commenting on actually states in its abstract: > Over a wide range of document types, Binoculars detects over 90% of generated samples from ChatGPT (and other LLMs) at a false positive rate of 0.01%, despite not being trained on any ChatGPT data.