This paper says they classify it by feeding tokens to a Google model.