How AI text watermarks hide inside word choices, not in the characters themselves: a visual, step-by-step explainer. Covers also how to erase marker. (James Padolsey)
- Text watermarks don't hide in characters or metadata, they live in word choice patterns, created by a secret key that subtly tilts the model's probability distribution toward a "green" set of words at each decision point.
- Detection is a statistical count, not a style judgment, and only the key-holder can run it, which means GPTZero and similar tools have nothing to do with this.
- Light paraphrasing dilutes the mark but doesn't erase it. What kills it is a full rewrite that shares no runs of wording with the original.
