We also less commonly see exposition that is not germane to a story, so a character is rarely even mentioned to be "weak", "intelligent", etc unless there is a point. And sometimes the point is that they are later shown to be "strong", "absent-minded", or other contradictions. Which means that mentioning a character's strength makes it more likely they will later be described as weak, than if it was never mentioned at all. Finally, double-contradiction is less common in human text (maybe because plain contradiction is sufficiently interesting), so a running text with no reversals is more likely to eventually reverse, than a running text with one reversal is to return to its original state.
While I don't agree at all with the author's sense that this represents some kind of "alignment" danger, it does go a long way to explaining why ChatGPT is easy to pull into conversations that shock or surprise, despite all the training. It's because human writing often attempts to shock and surprise, and the LLM is training on that statistically.