On average the distribution of selected watermarked tokens is the same as the original distribution. You can test this experimentally.

I think there is a possible weakness in the context of the watermarker but that is not your claim iiuc.