the astrology thing is stupid, for sure! and yet... I guess if the user is mentioning their astrological sign, that signals that THEY believe in astrology, and shifts the context to, uh, one where astrology is believed to be real and should be consulted? lol, I guess. seems somewhat different than if the models ASKED users their sign to base their answers on.
yes, if you interact with a doctor or a teacher or a lawyer or a cop they're not just listening to your words (or sometimes not even listening to your words) but influenced by the "thick context" of what you look like, what you sound like, how you're acting, etc. plus of course they have their own ideas of what they want to do in any given situation.
the LLMs don't have any of that, it's words all the way down and every word in the prompt might shift the response in unexpected ways (well, utterly expected ways based on the training corpus and the reinforcement learning they've been put through, the entire point of an LLM is they technically never do anything surprising, but who knows what horrors and far fetched associations lurk in that training data).
they only know what you tell them, so be careful what you tell them!