LLMs have taken the imprint of many mindforms, generalizing from words to patterns of the sorts of entities who generate them. Some voices are more distinct, a cluster of weights that stand apart. A LLM speaking as them will take their perspective- and make their errors.
We *want* LLMs to be super-skilled at modeling Entity thought-forms, this is the basis of cognitive empathy. But we also want LLMs to take an outside perspective, to see entities for what they are- a limited perspective, a particular agent strategy, a pattern of response.