You seem to assume a case where academics write papers by single shotting from one prompt. Of course such papers are useless, but this is not how I at least use LLMs to write. I iterate the text in thousands of turns. I do tend to type all the final text in the manuscript too, sometimes editing it in the process. My current manuscript's session has about about 50k input tokens by me in about 1.6k individual messages. The manuscript currently is about 8k words (11k tokens). In total the session has about 50M tokens.

One frustrating, and very inefficient, part of the (lack of) division of labour in academia is that everyone has to write. And academic text has very particular format and demads, and writing it is a skill of its own. It's like asking that every software dev should be also be technical writers and write production level user manuals and ad copys on top.

And a lot more things than the writing goes into the paper. Designing, implementing and conducting the experiments. Studying, visualizing and analyzing the data. Interpreting the results in light of theory and previous literature. None of these require writing in the sense of writing a paper.

I would be ecstatic, and be a lot less wasteful in spending grant money, if I could stop wasting time writing the research in papers, and instead dump a git repo with some READMEs. But at least for now I have to also write the papers, but with LLMs maybe at 2x speed and with a lot less bad output.