LLMs seem to do more of what brains do unconsciously, rather than consciously. Which means brains may be better at rating e.g. trustworthiness of some text, but they don't surface specific ratings to the conscious level. Meanwhile, language models seem to be able to expose those features as knobs, allowing you to boost or attenuate them. So you get to drag the e.g. "excited" slider down to minimum, and get a text that may be easier to process at a conscious level. Having a slider to remove rhetoric from text would be really useful development.