My honest take: Requiring that LLMs produce watermarked prose is not only wrong — it's absurd. That the generated text is unique and correct is a load-bearing feature of the technology. The word choices are what carries the meaning. Different words, same principle — before watermarks LLMs were entirely undetectable. It's not rocket science; it's a tool.
krig GoToSocial Actor feed
My honest take: Requiring that LLMs produce watermarked prose is not only wrong
Comments
Landa@graz.social 3 days ago @krig I‘m not sure how „generated text is unique and correct“ is a fundamental feature of LLMs?
The former can’t be because there’s no uniqueness check and there’s huge number of prompts that actually require identical answers, and the latter because, well the error class commonly described as „hallucinations“ seems to be a perfect counterexample.