It's not hard to understand. People already trust the output of LLMs way too much because it sounds reasonable. On further inspection often it turns out to be bullshit. So LLMs increase the level of bullshit compared to the input data. Repeat a few times and the problem becomes more and more obvious.