Who are LLMs simulating when they read?

Different people think different things are 'brave' or 'stupid'. But if we take multiple actions X_1, X_2, ..., X_N then we will not be able to find an ordering where everyone who thinks X_i is stupid would think X_i+k is stupid for k>0. Thus, to simulate people bravery reactions, LLMs must have a multidimensional space to capture it. Do they simulate bravery while reading a narrative or do they figure out bravery on demand when asked? I would wager it's the previous one, mostly, which raises the question: who's standards of bravery do LLMs simulate while reading a story? I doubt it's a thick enough simulation to capture everyone.