AI made code cheap to write. Comprehension is still stuck at human speed.
A sharp piece making the rounds nails what most teams miss: AI didn’t make engineering faster, it just moved the bottleneck. Writing code got cheap — output per engineer reportedly jumped 60% in a year — but reviewing it didn’t. One developer can now open six pull requests a day; whoever reviews them still has the same hours they always did.
And it’s right, completely. Review used to keep up easily — there wasn’t much to review. Now the queue never drains. And the job itself changed: you’re no longer mostly checking “is this correct,” you’re asking “should this code exist, and will we want to maintain it?” By one survey, 96% of developers don’t fully trust AI-generated code, and AI-written code reportedly surfaces 1.7× more issues than human-written. That’s harder work per PR, right as the pile explodes.
But the usual fixes — triage, load-balancing, smaller PRs, automating the mechanical checks — treat the symptom. They make the queue move faster. They do nothing for the real constraint, which isn’t review. It’s understanding.
We pushed creation to machine speed; comprehension is still stuck at human speed, and no amount of queue-routing changes that. When code arrives faster than anyone can build a model of it, “approving” a PR quietly shifts from “I understand this” to “it looked fine.” Speed that up and all you’ve done is merge code no one grasps.
A faster queue just gets you to the mess sooner.
