The AI Code Review Bottleneck

Every delivery pipeline has exactly one tightest constraint. For decades it was writing the code, and the whole discipline optimized around that: reviews were quick because diffs were small, authored by a colleague whose judgment you knew, in a context you shared. AI generation removed the writing constraint – and the constraint did what constraints do: it moved downstream to the next scarcest resource, human reading and judgment.

What makes this bottleneck harsher than a normal capacity squeeze: the new volume is also harder per unit. 38% of developers say reviewing AI code takes more effort than reviewing a colleague’s – there is no author to ask, no shared intent to lean on, and the failure modes are subtler.

The METR result deserves its place in this table: in a randomized trial, experienced open-source developers were 19% slower with AI assistance on familiar codebases – while estimating they had been faster. Perceived speed and delivered speed diverge, and the difference hides in exactly the verification work this page is about.

• Rubber-stamping. Approvals keep flowing at the old pace against the new volume – which is only possible by not really reading. The review exists as ritual; correctness ships on trust. This is verification debt accumulating at maximum rate with a green checkmark on it.
• Queue collapse. Reviews stay thorough, so they queue. PRs wait days, authors batch more changes into each one (+51% PR size is partly this), and bigger PRs are harder to review – a feedback loop that ends with the team slower than before AI.

Most teams oscillate between both, which is why throughput metrics alone look fine while debt metrics deteriorate.

You cannot hire past a constraint that scales with tooling. The relief comes from changing what reaches the reviewer:

1. Written intent per task – so verification has a reference and the reviewer stops reconstructing requirements from diffs (checkable specifications).
2. A machine pass before the human pass – types, tests, boundaries, and the spec comparison run per change in CI, so human minutes buy judgment instead of mechanics (two-pass workflow).
3. Evidence attached to every change – the reviewer reads what was verified instead of re-deriving it.

Reality Graph attacks the bottleneck at its cause: it makes each change arrive pre-verified – checked against written intent and boundaries, with validation the model did not author and an evidence report attached – so human review shrinks back to the judgment work only humans can do.