Sorry you're getting downvoted, this interpretation doesn't seem that far fetched to me. Here's the strawman: The text-based watermarking is going to be done p

Sorry you're getting downvoted, this interpretation doesn't seem that far fetched to me. Here's the strawman: The text-based watermarking is going to be done procedurally instead of generatively. Maybe they add some sequence of zero-width Unicode characters to all generated text at certain intervals. Then, there is effectively no false positive possible (because humans would [effectively] never type such sequences of unicode naturally). It may survive some editing (depending on how you select/edit the characters), and it's possible to be stripped (false negatives).