the bit about rereading the same function at midnight while the tests are green is painfully relatable. i think the problem is we tend to trust the output because it looks confident, not because we understand it. treating AI-generated code as a first draft you have to actually read before merging is the only way i have found to stay sane.