You seem to be handwaving all concerns about the actual tech, but I think the fact that "training" is literally just plagiarism, and the absolutely bonkers energy costs for doing so, do squarely position LLMs as doing more harm than good in most cases.
The innocent tech here is the concept of the neural net itself, but unless they're being trained on a constrained corpus of data and then used to analyze that or analogous data in a responsible and limited fashion then I think it's somewhere on a spectrum between "irresponsible" and "actually evil".