I don‘t know if having a bunch of unit tests is the way to go.

In my experience the AI will touch eberything. It is teached to solve problems with more code then less code. It will refactor your whole service for a simple change and because most of the unit tests are white box tests with poor refactoring stability they have also to be changend, and then i can trust them anymore.

I think what will help is more Integration Tests with harder interfaces, so the AI can change the code under it and i can more trust the Integration Tests

We desperately _want_ tests to prove our code is correct. Unfortunately, this is not possible. Instead, thinking of tests as a model is the way to make it an effective use of your time.