Consider they're trained to respond to the prompt you entered. If you enter "I'm in Czechia. Please tell me a local LLM model." of course you expect it to say something about Czechia since you queried it that way. Now imagine the first sentence is in the system prompt instead - same thing.
Comments
Another curious behavior here is: LLMs are eager to use information available to them, even when it's completely irrelevant.
Consider they're trained to respond to the prompt you entered. If you enter "I'm in Czechia. Please tell me a local LLM model." of course you expect it to say something about Czechia since you queried it that way. Now imagine the first sentence is in the system prompt instead - same thing.