A closer look at the signals our classifier uses to identify generated text with a 0.7% false-positive rate.