I finally read and understood the SynthID paper and now I'm against LLM watermarking. It is only non-distortionary in a fairly weak way

To give a concrete example if an llm predicts the next token will be a or b with 50-50 probability and our watermarking function g gives 1 to a and 0 to b we'll sample a 75% of the time

I feel like the proponents have been eliding this point and implying it only changes the random seed. Which I would be okay with. But unfortunately doesn't let you predict if a given snippet was in an llm's output because it's not robust to crops/edits

I do concede that in practice humans can't distinguish the text so I'm not strongly against watermarking. But I don't think that's the standard by which this should be judged

I also think watermarking is not a very useful technology relative to pangram style detectors because there are a lot of llms including open ones