This is a good experiment, but important caveat on p. 18. The effect is not because LLMs have some fiendishly clever manipulation strategy. It's mainly that they're fast, fluent, and patient — and these are effects that wear off a bit as people get used to them.