Clean code from an AI is not proof it works. Human bugs cluster near visible hesitation: awkward conditionals, vague variable names, an unfinished comment. AI bugs cluster somewhere else entirely, and reviewers trained on human tells walk right past them.
The Details:
Human developers leave traces of uncertainty in their code because they actually feel uncertain. A model does not experience doubt, so its output looks equally polished whether the underlying logic is verified or guessed. That uniform fluency removes the exact signal reviewers rely on to spot trouble.
The failures concentrate at the seams where generated code touches the outside world: library versions, API response shapes, sentinel values. A pagination loop that checks for a null token will run forever against an API that actually returns an empty string, because the model pattern-matched from other APIs instead of checking this one. Nothing in the code hints at the mismatch because the model never registered doubt about it.
Two related traps follow the same pattern. One is the plausible import: a function name that sounds exactly right but was renamed two versions ago or never existed. The other is silent error handling, a try/except block that looks defensive but quietly swallows the exception that would have exposed the real bug. Both read as competent because they were written with the same unearned confidence.
The fix is a new question during review. Instead of asking whether code looks right, ask what would have to be true for it to work and how that gets confirmed. Two concrete habits help: ask the model to list every unverified assumption it made, since it will admit them directly if prompted, and run the code against real APIs and real data earlier than usual, since AI output passes a reading test far more easily than a running one.
Bottom Line: AI code should not be trusted less than human code, it should be trusted for different reasons, with verification aimed at confidence rather than mess.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.
