I don't think the universal applicability of Sturgeon's Law necessarily leads to a contradiction here.

My incredibly, almost insultingly oversimplified model for what I, a lifelong consumer of text, feels is happening here goes like:

90% of culture is "bad". 10% is "good". Previously, we'd iterate on the 10% and (I remind you: oversimplified) discard the 90%, but `derive :: culture -> culture` still follows Sturgeon's Law, so at the next iteration, 90% is "bad" and 10% is good. But now what's happening is we're iterating on the whole sample, and a polished turd is still feculent. So in the next iteration, 90% + 90% of 10% of the culture is "bad". And so on.