> From a CS perspective I can't think of any way to have a watermark that couldn't be easily defeated through additional processing
The way that the current ones work is that they tweak the probabilities to generate specific patterns in the output. Where there's an equal probability of two words following another in the raw weights, they'll tune it so that there's a higher probability of one than the other.
If you know the weights and know the biassing factor, you can look at each word pair and see what the probability would be of the model generating that.
This means that the watermark is smeared all over the output. And it's not a binary thing though, each pair of word contributes something to the probability of matching the watermark and looking at the whole thing will give you a probability at the end.
Changing every other word should give you close to a 0% of matching the watermark but, at that point, why bother using the LLM at all? If you're going to change half of the words, you may as well just write them yourself. And that's a problem for people who want to share low-effort slop and pretend to be creative.
