Hi! We have some details on potential attacks in practice in Section 7 and discuss how hard it would be to brute-force the watermark in Sec. 5, at least to our

Hi! We have some details on potential attacks in practice in Section 7 and discuss how hard it would be to brute-force the watermark in Sec. 5, at least to our current knowledge. One important thing though: By design this watermark has a low false-positive rate (human text flagged as machine generated). Human text is in expectation 50% whitelisted (because the watermark PRNG is unknown to the human writer) with a clearly defined variance. We can then carefully set up a hypothesis test to check for machine-generated text (see ROC curves in Fig.4), with tiny FPRs.