There are many, many reasons to be skeptical of AI. There are also excellent tasks it can efficiently help with.
I wrote a project where I'd initially hardcoded a menu hierarchy into its Rust. I wanted to pull that out into a config file so it could be altered, localized, etc without users having it and recompile the source. I opened a “menu.yaml” file, typed the name of the top-level menu, paused for a moment to sip coffee, and Zed popped up a suggested completion of the file which was syntactically correct and perfect for use as-is.
I honestly expected I’d spend an hour mechanically translating Rust to YAML and debugging the mistakes. It actually took about 10 seconds.
It’s also been freaking brilliant for writing docstrings explaining what the code I just manually wrote does.
I don't want to use AI to write my code, any more than I'd want it to solve my crossword. I sure like having it help with the repetitive gruntwork and boilerplate.
This sort of extremely narrow use case is what I think AI is good for but the problem is that if you have it for this one you will use it for other things and slowly atrophy.