I am not so skeptical about AI usage for paper writing as the paper will be often public days after anyways (pre-print servers such as arXiv).

So yes, you use it to write the paper but soon it is public knowledge anyway.

I am not sure if there is much to learn from the draft of the authors.

I think the goal is to capture high quality training data to eventually create an automated research product. I could see the value of having drafts, comments, and collaboration discussions as a pattern to train the LLMs to emulate.

Why do you think these points would make the usage dangerous?

They have to monetize somehow...