[arijun]

I disagree. There is more content out there than I can read in many lifetimes, so I have to be selective. LLM generated text (like any text) can be well put together on the surface level but require deeper consideration to see the flaws, and of course this takes more effort than the writing did.

A human-written piece indicates someone believes in it enough to put in enough effort to write it up nicely, so it works as a heuristic of underlying quality.

[rdiddly]

All true, but how do you distinguish human-written from AI-written or a hybrid? They all have an author's name attached. You would have to limit your reading to people you know personally. (Which isn't a terrible idea actually.) Otherwise it's a judgment call, which inevitably comes down to a question of writing quality. "This has to be AI because it's so terrible." But humans are perfectly capable of writing terribly (that is in fact where the LLM learned it) and LLMs can even write well occasionally, including with human intervention. So I decided that if I'm using quality as a proxy to guess at authorship, why not just forget authorship and make quality primary. Basically since authorship is unknowable I'm declaring it irrelevant. It's not ideal but these are the times we're living in.