if the AI breaks something it will be apparent.

No, it isn't usually apparent. Most bugs from AI are subtle edge cases that will only show up later. One of those "later" instances night be when you try to run it on a different computer, or when another user presses buttons in an order you didn't try, or when you try to add a new feature, or worse, when someone tries to hack your code and finds all the common security vulnerabilities because that's what the AI has seen most often and copied.

AI can't think – it can only give results it has seen before or some combination of them. We already know AI can't produce an image of a full glass of wine, or tell you what year is next year on Jan 1st.

The technology can't work. No matter the input prompt, it just isn't capable of producing code with no bugs or edge cases. That requires thinking and logical deductions, both disproven by the examples in the previous paragraph.