You know why the code is wrong because you have the experience to see where the issue is and what it is.

If you've learned coding with LLMs from the start, you won't acquire the experience needed to be able to tell what is wrong.

I've worked with a client that tried to generate code for a HCI bluetooth device, trying to recreate the full Bluetooth stack, instead of picking the right product from the start, with a working stack.

And that was a client that had technical knowledge, just not for Bluetooth and HCI.

And if you try to tell the AI what's wrong, it will create bullshit code until it kinda works, adding more issues along the way.

I'm sure that AI will replace coders one day, but LLMs aren't AI and they are neat ready to write decent, complex code.