On my last job before retirement, we were asked to explore using LLMs to extract information from reports intel analysts have carefully prepared. I noted that the LLM summaries, even when factually correct (it made up quite a bit of stuff), often missed the judgment conveyed via nuanced words chosen by the human analysts. I'm sure nobody read my report, just like they couldn't be bothered to read the reports by trained analysts and they wanted a shorter version from a stochastic parrot instead