Yours is not the first post asking about this here. Which in and of itself says something.
I don't use them, at all. I briefly tried the local tab completion stuff offered in JetBrains products. It lasted an hour or two. The log messages it wrote didn't sound like me, and the "copilot pause" was immediately frustrating.
The boilerplate argument comes up a lot, but I really don't see it as the huge issue that would drive me to try and make Clippy generate it for me. That sort of "boring" work is great for "meditating" on the thing you're doing. Spending time adjacent to the problem putting up the scaffolding makes you mentally examine the places where things are going to interact and gives that little seed of an idea time to grow a bit. Become more robust.
Later, when there's an issue, you can ask the human that wrote something questions about it and they will probably have at least a fuzzy recollection of how it was built (and why it was done that way) that can offer ideas. Best you can do is hope the LLM doesn't hallucinate when you ask it about all the broken stuff.
Ultimately I see neither value nor "power" in the current "assistants." They generate the statistically most median output and often get it wrong. They make stuff up. They have no understanding of anything, and they don't learn from mistakes. If they were a person you'd be asking serious, but nearly rhetorical, questions about whether or not to fire them.