The author's relevant experience is in developing Copilot, not in information theory or statistics. So, when they say:

Unlike image and video watermarks, text watermarks will always be trivial to remove.

This is wrong; image and video watermarks are also trivial to destroy. In general, watermarking only works for physical objects and fails for information-theoretic objects.

Watermarks fail. PersonalIP survives.

Perfect is the enemy of good. The worst offenders for AI slop are the sort of people who won't bother making the extra effort to scrub watermarks.

Sorry, I couldn't make that out. Maybe you could remove Altman's cock from your mouth first?