It's like if on any website you went to you saw a lot of posts written by the same guy over and over again. Even if he used different names, you'd start to recognize him eventually because of his style. Seeing as he doesn't say a lot of valuable stuff, you'd also learn to skip whatever he says.

I do worry that it's just survivorship bias and we're also consuming higher-quality AI output that's indistinguishable from human writing, but we focus on the raw, unedited, low-effort AI slop and think that we're good at recognizing AI text. Even if we really are at the moment, it might not be long until AI companies figure it out. I'm not sure why they haven't yet, given how many books they've burned for this already. Maybe it's just more efficient for the model to stick to a single way of writing, I don't know. But when that point comes, we'll be back to the usual way of reading and interpreting text because there would be no way to tell what produced it.