I don't think the claim will bear out that LLM-generated text fares more poorly on comprehensibility tests than those of the median literate speaker. I understand the aesthetic arguments against it, and largely agree, but the technical argument you're making here is almost certainly and starkly false.
Most people write quite badly!