AI writes the code. The review is still yours.
AI-assisted development moved the bottleneck from writing code to judging it. That changes what a senior engineer is for, and it makes review skills more valuable, not less.
- AI-assisted development
- Code review
- Clean code
I use AI assistants every day. They are genuinely good: boilerplate disappears, unfamiliar APIs become approachable, and a working first draft of most features costs minutes instead of hours. None of that is the interesting part. The interesting part is what happens after the draft exists.
The bottleneck moved
Before assistants, the slow step was producing the implementation. Now the slow step is deciding whether the implementation should survive: whether the abstraction fits the system, whether the query pattern scales past fixture data, whether the error paths exist, whether the code belongs at this layer at all. Generation is cheap. Judgment is not.
This is not an argument against the tools. It is an observation about where the engineering now lives. A model optimizes for plausible: code that compiles, reads cleanly and matches the shape of its training data. Production systems fail on exactly the dimensions plausibility does not cover: load, concurrency, partial failure, the requirement that arrives next quarter.
What review means now
I review generated code the way I review any pull request, with the same first question: what would make this wrong? Not "does it work". It usually works. Wrong means an N+1 hiding behind an ORM call, a client component doing a server's job, an any that erases a contract, an abstraction built for a future that will never come. These are architectural judgments, and they are invisible to anyone, human or not, who only checks whether the feature functions.
The practical discipline is refusing to rationalize. When generated code is close but structurally wrong, the temptation is to keep it because it runs. Replace it. The ten minutes you save by keeping a bad structure are borrowed at interest from every engineer who touches the file after you.
Faster and better are not the same axis
Teams that treat AI output as finished work accumulate debt at generation speed, which is a new and impressive way to accumulate debt. Teams that pair generation with real review get both halves: the speed of the tools and the durability of engineered systems. The second half is a human skill, and right now it is the scarcer one.