And if the tests change, does it have to fully rewrite the code via AI?

If you had code that was reasonably working before, but you change the tests and the AI does the "full rewrite" for you, how do you have confidence that the code works? Go ahead, say "it passed the test suite".

One of the things that happens in software dev is the devs grokking the overall desire/aims of the software and its users, and often anticipating things that aren't even in the requirements.

The test suites are approximations of validating the requirements / derived intent.

But what faith do you have in the AI black box, or the black boxes it generates?