AI is getting good at reviewing code.
That is exactly why the human's job in review has to move up.
I have watched this shift happen in practice. An AI reviewer is genuinely good at the mechanical pass: syntax, common patterns, security scans, "this function does not handle null." On a large diff it reduces the line-by-line burden a human used to carry. That part is real, and pretending otherwise is nostalgia.
But there is a failure mode hiding inside that usefulness. A bot will happily leave thirty comments about spacing and miss the one change that quietly breaks production, because the thing that makes a change dangerous is rarely visible in the diff. It is whether the change fits the architecture, whether it respects a boundary that exists for a reason, whether the business logic is actually correct for this domain. Should a 15% discount apply to premium users? Is this the right abstraction, or just a working one? Those are not pattern-matching questions, and the model answers them either generically or confidently wrong.
So the reviewer's value moves up the abstraction stack. Not "where is this called from" — the AI can find that. The human supplies what the AI structurally cannot: recent decisions, architectural principles, ownership knowledge, why a past simplification is forbidden, and accountability for accepting the change. The danger is the opposite reflex — reading the AI summary and clicking accept. That is not review moving up; that is review disappearing and calling itself faster.
The mechanical reviewer can be automated. The reviewer who owns the consequence of the change cannot.