Hmm well people who write code and I know a bunch of them who support AI and I support them in my line of work mentioned this very recent survey to me from late last year.
A comprehensive survey of 1149 developers reveals that nearly half of programmers consider AI-generated code to be functionally incorrect, with only 4% trusting its correctness completely. The data exposes a growing gap between AI hype and developer reality.
I work with AI quite a bit. My experience is that I was able to get Gemini to write me an AI agent correctly, but it took many iterations, with me consistently refining the requirements and working through bugs. I also got our in-house, sandboxed work AI to write a full-on, dependency-injection version of a method I wrote, with mocks and automated unit tests, with very little by way of the iterative process.
So it can work.
But you need to use prompt frameworks for the best results. I use TACO: TAsk / COntext.