A theory of LLM output and negativity
One way to tell you're reading LLM output is that it doesn't really go anywhere. There might be a string of observations, and part of it might read as a conclusion, but there's unlikely to be a substantively progressing argument.
Consider this just as a structure: it's punchy at the micro level but is limited to repetition or other simple formulas at the macro level. What kind of content suits that structure best?
I suspect that it's condemnatory, fear-mongering, and neurotic content. This is true quite apart from AI, and shows up in other domains. An effective condemnatory speech might consist of thinly connected criticisms; it tends not to need a satisfying, non-obvious conclusion. A speech of praise, conversely, tends to bore us if it has an analogous structure. Sympathetic biographies are impressive largely because they slowly teach us about someone and lead to a non-obvious conclusion about their life and character; those are rare. Diss tracks and roasts work pretty well; "praise tracks" tend not even to try to go anywhere ("Kind and Generous" and "Praise You" are the first two to come to mind, and they're both repetitive in a strikingly non-varied way). Commentators disagree on how well it worked when Clint Eastwood yelled at an empty chair for twelve minutes, but it's hard to imagine someone even trying to praise someone by directing loosely-connected compliments at a chair for twelve minutes.
Limericks are good for jokes; the structure suits the content. LLM output, if I'm right, suits negative feelings better than positive ones, again for structural reasons that are independent of training data, sinister corporate motives, or whatever else.
Two final notes:
- This post started life as an attempt to understand a phenomenon that struck me as strange: why are serious, thoughtful people approvingly citing articles warning us about cognitive debt when those articles are are obviously AI-generated? I'm sure that there's plenty of LLM slop in breathless support of coding with AI, but it's hard for me to imagine such output being satisfying enough to endorse.
- I think I've underestimated this phenomenon as a driver of "ragebait" and other ways that social media can skew negative. We might enjoy streams of cat videos and sports highlights, but I don't think we perceive them as combining to form a significant whole. Conversely, people do seem to find cumulative significance in piles of unrelated media with negative affect. The structure just suits negativity better.