Review as a practice carries a hidden assumption, which is that writing is slower than reading. A developer spent two days on a change, so spending forty minutes reading it felt proportionate. That ratio held for as long as code came out of a human at human speed, and it silently underwrote every quality gate we built on top of it.
Generation is now effectively free, but comprehension costs exactly what it always did. The ratio has inverted, and no process was updated to reflect it.
What happens to review under AI-assisted development ?
It degrades quietly rather than visibly. Reviewers still open the diff, still leave comments, and still approve, but the depth of the read drops as volume rises, because attention does not scale the way generation does. Approval starts to mean "nothing looked wrong" instead of "I understand this."
That drift is where most AI technical debt originates, and notably it does not show up in any metric you currently track. Delivery speeds up, review times drop, and the debt accumulates underneath perfectly healthy numbers.
Debt now has a distribution mechanism
The older, slower version of this problem stayed local, since a bad pattern lived in the file where someone wrote it until a human copied it somewhere else. That was the natural rate limit.
Assistants removed it. A pattern in the codebase is a pattern the model reads as intent, and it will reproduce that pattern everywhere it seems to fit, without ever distinguishing a deliberate design decision from a random compromise. As a result, a single shortcut stops being a shortcut and becomes a convention, usually within a sprint.
Comprehension is the gate, not correctness
Review the pattern rather than the diff, because the diff is no longer the unit of risk. Make the reviewer state what the change does in one sentence before approving, since a reviewer who cannot do that has not reviewed anything. Track how often you generate code in areas nobody on the team can currently explain, and treat that as a real signal.
The uncomfortable part is that none of this is a tooling gap. The tools work. What has not kept pace is the assumption that somebody, somewhere, still understands the system.
Top comments (0)