AI for writing feels like such a stupid idea.

When you write things you just need to think about what you want to say and write it. It's not hard.

If you make an LLM generate the text, and you still bear responsibility for whether it's correct or not, you haven't actually reduced the amount of work you have to do at all because now instead of "thinking -> writing" you need "reading -> thinking -> rewriting."

"Reading is easier than writing" sounds like something that someone who never writes anything would think. In both code and natural languages writing is easier than reading. Text is after all a representation of thought, so to write you just need to represent your thoughts in a way others can understand, but in order to read you must decipher what the author was trying to convey from their written word. If the author is an LLM that is going to be a lot more difficult and annoying.

I've had similar feelings about LLM's "completing" code. It "feels" like I'm more productive, but at the same time the "completions" feel more like constant interjections that won't let me get into a flow of writing, and often when things don't work it's because the AI quickly generated code that looked correct at first glance, so I left it that way, but when I tried to run the program it turns out it was wrong, and when it happens I can't help but think: of course it's wrong, I haven't written the code yet. There is code, the LLM wrote it. But I haven't written it. I haven't done the work. So it's like having a "placeholder" that you still HAVE to check but there is no metadata that says this is a WIP, no # TODO: check the LLM output. It's a placeholder camouflaged as finished work.