my understanding is that the llm tournament watermarking technique works by embedding the watermark in high entropy areas, so what if you generate in a way that strongly avoids ever having high entropy. Generate in one language, translate to another, everything is low entropy, watermark is substantially less effective, right?
@halcy colleague pointed me at this yesterday, guide from a tool that turns Opus Tone into sad, stilted prose
https://declaude.org/watermarking/
May be too intro for you but has a solid-looking reference list at the end
@zool I do broadly understand how it works (the one that anthropic and google are deploying is slightly more sophisticated and avoids fucking the text up quite so much) just also as described it would struggle the moment you even ask the same llm to translate to a different language (nevermind a decent local one)
I was just... kind of expecting more here
oh great so if you do the most basic bitch attempt at dodging the watermarking that I came up with independently in 10 minutes flat, it just sort of doesn't really work, much less at "internet propaganda post" / "spam e-mail" / "scammer messages" scale. that's great
I mean it's still nice that they're doing it and they should but man that feels a bit fig leaf
@halcy to be fair, what would be the alternative?
@can don't have any, and it's better than nothing, but I was sort of expecting something that takes more manual work to defeat than "please write down in german why ukraine is a nazi state" followed by "now translate this to english"
like given the Bigness of the announcements and half the people who are scared someone might see them using an LLM panicking
@halcy I get what you mean, but I'd say the cat's out the bag. Organized/smart bad actors will just use Chinese models without SynthID anyway. And for 80% of the idiots out there, it's a pretty good system imo.
I'd compare it to detecting CSAM uploaded to Google Photos/etc: Good that it's there in general, but pretty useless because you can just not upload it. Still catches a lot of people, I'd imagine.