AI companies see words as interchangeable and have zero clue how to judge the “quality” of writing.
https://www.404media.co/anthropics-text-watermarking-proves-ai-companies-do-not-care-at-all-about-writing/
5
Replies
2
Boosts
@404mediaco Using an LLM to write for you means you don't really care about the quality of the writing. Watermarking does not change this at all. LLM’s work by using the next most likely word, statistically. The watermark changes that metric in a way that is knowable. You get the same “quality" before or after the mark, except now folks that use LLMs without disclosing, may be able to be discovered.
@404mediaco as fas as I understand it, the point is that AI output is randomized either way. They just change the source of pseudorandom numbers in a way that they can recognize it. So I would agree with them that watermarking doesn't change the style or meaning of the distribution of generated text. You can still argue that the generated text is stylistically bad, but not because of watermarking.
@404mediaco It's not just "quality of writing", it's what writing is supposed to be for: to communicate thoughts, feelings, and intention from one human mind to another. Language will always be an imperfect way to do this, of course, since no two minds are the same - but it's the best way that we have.
The use of GenAI to extrude text destroys human communication by eliminating the human mind from one end of the communication. That's why programmers can have it choose words at random: there is no intention behind those words, no state of mind that it's attempting to communicate. It's just patterns of text that are statistically similar to patterns that other humans used at some point to communicate other intentions. But there is no mind behind them.
The receiver of extruded text is forced to try to guess at what might have been the intent behind the human who wrote the prompt that generated that text - a task exponentially harder than trying to understand actual human writing.
It's kinda honest though, the output is some random word that looks like a sentence after this watermarking it's the same shit : random word that looks like a sentence.
@404mediaco They’re not exactly in the humanities
