It works for them not scraping their own slop back into training data. I assume that is actually the real purpose of the system. They don’t want to share the key with the public. But they probably will with other llm companies in exchange for theirs.
My tinfoil head tells me they don’t want to ingest their own slop for training. But this only helps them if others do the same. Data poisoning is a real problem for llm future.
This won’t work - you can always use a dumber, open model to rephrase and remove the watermark. All this does is marketing for anthropic, which I see most people are buying. Gives you a false impression you will easily be able to figure out if it’s done by AI, which it is not. Liars are still gonna lie, cheaters still gonna cheat.
Endgame is training data is audited and ethically or at least legally acquired for ALL companies making models.
Yeah, you know that episode where Homer does pranks for Mr Burns and goes in and buys a no 1 spiderman from the comic book guy and eats it in front of him. That’s how most people feel about book destruction. Nothing legally wrong, but disgusting nonetheless.