Why is AI so bad at fixing bad back-end?

AI definitely has come a long way, progressed beautifully in front-end, but when it comes to fixing bad backend code, it still struggles a lot.
The complexity often requires human intuition and deep understanding.

It can help spot some issues or suggest improvements, but relying solely on AI to untangle messy backend problems often falls short.

Has anyone had surprisingly good or bad experiences with AI in this area?
I'd love some advice (or you're free to rant)