Idk, it might seem obvious, but I noticed something interesting when I was thinking about the "it's not x, it's y". Like a friend told me some time ago, "it looks like infomercials, it's like it's trying to sell you something".
When someone says something like that, what comes to mind? My first thought is always "but I didn't think, say or reason that it would be x, nor you". Why cite a different thing before the real argument before the argument then?
To me this just seems some way to add "weight" to an argument, like a high school essay, taking attention out of the real argument, stylistically giving the illusion of strength. It looks like and acts like the high school essay "padding" that we all know and used (or at least people used at my school) "that guy was not important, he was essential", "the trouser gave the world one of the most important and necessary inventions", "the onion is one of the most notorious newspapers in existence, being cited in nature and by the bible itself".
I'm giving some bullshit texts with no meaning examples, but just substituting this with the "guy" we needed to talk about in 4 pages would add shit up and give a little bit of the semblance of weight to the text.
Now looking at any Claude or ChatGPT text, this is exactly the thing that's happening, a high schooler is padding their essay with things saying that "yeah, my argument is very strong, because it's not x, it's y", and people are falling for it. Every topic on the signs of LLM wikipedia page is this.
Idk, but this seems to be the same case of the "warmth" that bootlicking LLMs have, the person looking at the screen gets manipulated (by the companies that train the LLMs to act this way) and likes it, even if not directly or knowingly.
Imagine what could be done with this kind of "fallacies"? Just changing peoples mind on topics they want to be changed? Instead of refusing to say anything, the LLM trainers could just train certain argument tactics and try to change peoples mind, if people is falling for the bullshit padding, why not tactically engineer peoples opinions?
If it's not doing this already.
A core training principle in any LLM is to be helpful. This tell is a creative way to self-validate a claim, even if it is weak. A common goal that is difficult to do is to retrain an LLM to prioritize being factually accurate vs. always finding some answer that makes the user happy. The reward system for that behavior is very deep and hard to break. AI could have been... rather was in the ANI forms, a decent tool until someone figured out how to make profit from hype, and now the damage is done.
100% agree.
Sometimes I think if there's any way to make a "most correct" LLM, if it's even possible. It seems those don't make money hahaha.
Seems that's usually the case. The original internet and world wide web wasn't a big money maker in its design, but boy that changed once corporations took over and people became the product.