I'm a dev and I think my experience is mostly similar to yours. Where AI seems to work well, is when the boundaries of the problem are very well defined. For example : "Take this C++ implementation of the LCS algorithm and convert it to JS". That would have taken me a few hours at least, but AI appears to have nailed it. However, anything where a large amount of context is needed and it starts to fail fast, and suggest absolutely insane things. I have turned of copilot on my IDE because it slows me down, but I will still ask questions to chatgpt when I have a specific problem I think it can help me with. I also will ask pointed questions when I review other dev's code, and my expectation is the author can explain why they wrote it.