The most dangerous thing about AI-generated code is how finished it looks.
Plausible is not correct
A model produces code that compiles, reads well and seems done. Then it fails quietly on the edge case nobody asked about. The polish hides the gap.
Build an adversary
My most useful experiment was a verification agent whose only job was to try to break the output and report what failed. On a real build it caught three silent bugs a quick human review had waved through.
Verification is a people problem too
Inside an organisation, the question is not just “is the code correct” but “do people trust it.” Visible checks, clear ownership and human oversight are what make AI output something a team will actually rely on.
Ship fast, but never ship unverified. The check is cheap; the silent failure is not.
