A lot of times the act of specifying test criteria prevents developers from accidentally vibe coding themselves into a bad implementation. You can then read the tests and verify that it does what you want it to. You can read the code!

I’m not saying that it’s all hunky dory, but you use AI for straight up test driven development to catch edge cases and correct sloppy implementations before they even get coded by your giant chaos machine.