My understanding of the watermark process is that it only kicks in when it's not the case that one particular token is overwhelmingly likely. In the case where many tokens are plausible, we already choose with PRNG -- the "watermark" is essentially a more easily reproduced PRNG, but it still looks like PRNG.