Reviewing AI-generated code needs its own checklist

AI-generated code tends to fail in a specific way: it looks confident and is often locally correct, but it can miss the surrounding system. It will happily call a deprecated internal API correctly or handle an edge case that does not apply to your data model.

The checklist I use is different from a normal PR review. I look harder at boundary conditions the model could not have known about, at whether error handling matches how the codebase actually surfaces failures, and at whether tests were written to match the implementation instead of the intended behavior. That last one happens more often than people think, and CI will still pass.

None of this is a reason to avoid AI-assisted engineering. It is a reason to review it with different expectations, because the failure modes are different.