One point I haven’t seen mentioned in the recent AI text watermarking discourse is that watermarking output seems like it would be quite useful for detection and rejection during the model training process to avoid training models on data that can pretty reliably be determined to be AI generated, like random web pages and articles and such. A clever way to minimize that spiral. It sure wouldn’t cover everything that’s generated, but it would let a nontrivial chunk be readily omitted, I imagine.