I think this article falters because it takes ChatGPT on its face, and it treats it as anthropomorphic (if these methods can convince an llm, think about how they impact real people) rather than (hey these models are deeply vulnerable to coercion and will spit out what you want them to if you try)