How do you review code created by an AI if you were not able to write that code yourself?

The last time I used Cursor, I gave it a simple expense tracker project to test out the limits of its capabilities. The program ran right away the first time, and that made me suspicious rather than pleased.

I found out I couldn’t explain what half the functions did in the file. In any case where something was broken, I lacked context and could only guess at what prompts had caused it.

Now I speak each change back out loud in layman’s terms before accepting it. If I can’t describe the function in a sentence, I ask the model to either simplify or add comments.

This is slower, but the code feels more like my own again. How does one even review something they could never have produced? Is it a process or an intuition?