What it is
The PR is a wall of generated diff. The description is a three-bullet AI-written summary. The reviewer clicks Approve in under a minute because there isn’t a realistic way to read it all — and there are six more reviews queued behind it. The bar that “Approve” used to clear has quietly collapsed to “the queue must keep moving.”
How it happens
AI-generated PRs are bigger, more frequent, and arrive with confident-sounding descriptions the author also didn’t write. The reviewer’s queue triples. The cost per careful review doesn’t change. The time budget shrinks. The default move becomes: skim the AI summary, scan a couple of key files, approve. The reviewer is effectively reading an AI-written summary of AI-written code — a closed loop with no human judgment in it anywhere.
The AI-era mechanism is double-sided. On the author side, AI inflates the size and frequency of PRs while producing authoritative-sounding descriptions that paper over decisions the author didn’t actually make. On the reviewer side, AI velocity has no matching review-velocity multiplier — reviewers are still humans reading code one diff at a time. The asymmetry is the whole pattern.
Why it’s dangerous
When Approve stops meaning “someone qualified made a judgment,” it stops meaning anything. The signal is gone, and so is the catch rate — bugs land, security holes land, dead code accumulates, contracts drift, regressions ship. The team still believes it has a review process; it actually has a queue of rubber stamps that the postmortem will eventually have to acknowledge.
The compounding piece is worse. When reviewers stop catching things, authors stop pre-checking. PR quality drops to match what review actually catches, which is nothing. The next AI session generates against this lower bar. The AI-era hinge: review never scaled with code generation, and the traditional answer (“more reviewers”) doesn’t work — senior reviewers are scarcer, not more abundant, and throwing junior reviewers at AI-generated diffs is how the rubber stamp becomes the official policy.
How to prevent it
The author has to do the work that makes review possible — not delegate it to the reviewer. That means: PRs sized to be reviewable in one sitting, descriptions that explain decisions (not changes), and a self-review the author can defend out loud. AI can draft the description; the author owns every claim in it. If the author can’t answer “why this approach?” about any non-trivial section, the PR isn’t ready, no matter how clean the diff looks.
Scale to stakes: a typo fix doesn’t need a self-review essay. A new module does. The friction signal is the moment a reviewer can’t tell what the PR is for after a careful read of the description — the PR is wrong, not the reviewer.
The serious team fix
Three things, reinforcing each other:
- Author-driven readability as a norm. PRs are scoped to one decision, not a batch of changes; descriptions explain decisions, not diffs; the author runs through their own change before requesting review and pre-empties the dumb questions. “I would not want to review this PR” is the rejection signal authors apply to themselves. The team agrees this is the bar out loud.
- AI as first-pass reviewer. A slash command, CI step, or PR-bot that runs an AI review focused on specific risks — security patterns, contract drift, dead code, regression-prone changes, design inconsistency, test/code coupling. The AI surfaces a prioritized list of findings; the human reviewer’s job becomes judgment on the surfaced candidates, not first-pass triage of twelve-hundred lines.
- CI that catches what human review can’t reliably catch anymore. Type checks, lint rules, contract tests against the API spec, dependency scans, regression tests against canary inputs, license checks. Anything mechanical gets enforced mechanically and blocks merge. Human review focuses on design, intent, and the judgment calls mechanical checks can’t see — which is where humans are still indispensable and where their attention is actually spent.
The shift is: Approve used to mean “somebody read this.” Now it has to mean “somebody made a judgment call,” and the mechanical part has to be the machine’s job. Otherwise the reviewer’s name on the merge is doing work it can’t actually do.