It’s not about it being counterproductive. It’s about correctness. If a tool produces a million lines of pure compilable gibberish unrelated to what you’re trying to do, from a pure lines of code perspective, that’d be a productive tool. But software development is more complicated than writing the most lines.

Now, I’m not saying that AI tools produce pure compilable gibberish, but they don’t reliably produce correct code either. So, they fall somewhere in the middle, and similarly to “driver assistance” technologies that half automate things but require constant supervision, it’s possible that the middle is the worst category for a tool to fall into.

Everywhere around AI tools there are asterisks about it not always producing correct results. The developer using the tool is ultimately responsible for the output of their own commits, but the tool itself shares in the blame because of its unreliable nature.