How do I catch when AI code and analysis is wrong?

You have shipped work you did not fully understand, and it worries you. Good. That worry is your judgment still awake. The danger is not using the tool; it is handing over your understanding along with the task. AI is fluent, and fluent is not the same as correct. It will give you a clean query built on a wrong assumption, or a model that looks right and quietly misses the point. The only guard against that is you understanding the thing well enough to smell when it is off. That means slowing down at the exact moment the tool tempts you to speed up. Make one rule for yourself: you do not ship what you cannot explain. Before your next analysis goes out, write two plain sentences on why the result makes sense and what would break it. If you cannot write them, you are not ready to ship. That single habit will catch most of what slips through.