The problem with using them for brainstorming is that you've no idea whether the things (I won't call them ideas) they return have any connection to reality, so you need to sanity check all of them before even considering them. By the time you've done that, you could probably have come up with them yourself, possibly with the aid of a small rubber duck to talk to. You also don't know what it doesn't 'know', so you don't know which areas it's not addressing. The worst part is that the quality of the LLM' response will be approximately the average of its training data, so you're not even getting any particularly good ideas from it.

Whilst I've not tried using an LLM for translation, I have tried using them to draft things like letters, but found them woefull lacking and rather pointless. By the time I've written an adaquately detailed prompt, read what it returns, corrected the most egregious errors, and rewritten it so it's not in particularly weird, bland tone that they inevitably return, I could have written the document myself and got a better result.