Author's experience with ChatGPT as a coworker mirrors my own with Copilot. It definitely helps with the white page problem when doing green field work. Its really great at generating config since there are so many examples. Its pretty good at predicting tests, especially once you have built up a few in a file. I bet most tests will be AI generated within 5 years, or companies will run AI analysis on code they do for coverage.

It's not very good at high level stuff. It will attempt to explain code, and generate very plausible explanations that often miss key edge cases or forget variables. You have to keep on eye on it to prevent it from missing things. Recently I was doing some basic offset sum math, and it would readily generate an expression for the summed offset that would forget a term or two.

Copilot you can also prompt using comments, which means I first express my wants in plain language, and then code. Its made me realize how much writing code is tied to my thought process, and trying to decouple them is hard. Also copilot can be a net negative. If you are thinking about a line, it may give you a completion. Now you have to think about the code in your head and the possibly wrong completion on the page. Like during pair programming, I just want Copilot to shut up so I can think.