Code Review When the Author Is a Machine

Code review was designed around a human author: someone who understood the change, made deliberate trade-offs, and could answer "why did you do it this way?" When the author is a model, every one of those assumptions breaks — and review has to adapt.

Review intent, not just syntax

AI-generated code is usually syntactically clean and stylistically consistent, which makes it *look* more trustworthy than it is. The old review reflexes — nitpicking naming, formatting, obvious bugs — matter less because the machine rarely fumbles those. What it fumbles is intent: does this actually solve the problem, handle the edge cases, and respect the constraints that weren't written down? Review has to move up the stack, from "is this correct syntax" to "is this the correct thing."

The reviewer becomes the bottleneck

If one engineer can now generate five PRs in an afternoon, the team's throughput is no longer bound by writing — it's bound by reviewing. That's not a problem to engineer away; it's the new center of gravity. The highest-leverage person on an AI-accelerated team is whoever can read a diff and tell, quickly and correctly, whether it's safe. Protect that capacity.

Make the machine review the machine

The scale problem has a partial answer: use AI to triage AI. A model can flag the obvious issues, surface the risky diffs, and let the human spend their limited, expensive attention on the seams that actually matter — security, concurrency, data integrity, irreversibility. The machine handles breadth; the human handles judgment. That division is what makes the whole thing sustainable.