AI has a strange ability to be brilliant and awful at almost the same time.
An image tool can create something that looks like a professional photograph and then quietly add an extra leg to the chair. A language model can work through a complicated commercial problem and then ignore one very clear instruction at the top of the page.
The mistake is assuming that because it was brilliant at the difficult bit, it must also be reliable at the easy bit. It is not always like normal software where the same input produces the same tidy behaviour every time.
The useful habit is to separate “impressive” from “safe to trust”. Let AI do the thing it is unusually good at, then put a simple check around the bit that actually matters.