[lukifer]

The fascinating paradox: there are clearly "tells" (slop-smells, like code-smells?) of LLM-generated text. We're all developing heuristics rapidly, which probably pass a Pepsi challenge 95+% of the time.

And yet: LLMs are writing entirely based on human input. Presumably there exists a great quantity of median representative text, some lowest-common denominator, of humans who write similarly to these heuristics.

(In particular: why are LLMs so fond of em-dashes, when I'm not sure I've ever seen them used in the wilds of the internet?)

[Inityx]

https://news.ycombinator.com/item?id=46273466