They don't follow orders very well because they don't follow orders at all. Even if they could recognize trickery they wouldn't "know" that they're supposed to reject it. If the model predicts that the "evil" response is more likely than the "good" one, that is what it will produce.
These are sequence predictors. The fact that they sort of talk like people doesn't mean they work anything like people.
These are sequence predictors. The fact that they sort of talk like people doesn't mean they work anything like people.