AI generates faster than humans can validate. That gap is the actual bottleneck.
Throughput went exponential. Verification stayed linear. Something has to give and it should not be the checking.
The weak signal nobody is charting: generation capacity is compounding and review capacity is a headcount.
Every organisation I have seen hits this at the same place — not "can we produce it" but "can we trust it fast enough to use it".
The forces
Token prices fell 84–99% since early 2023 depending how you index. Context windows went from about 2,048 tokens in 2020 to a million as standard. Both curves point at more output per unit of effort.
Human review does not have a curve. It has a salary.
The tempting answer, and why it is only half right
The obvious move is AI for AI — use models to check models. It genuinely works for triage, and refusing it entirely means the bottleneck wins.
The trap is that a model checking a model inherits every failure mode of the thing it is checking. It is sycophantic. It is position-biased. Asked to verify, it confirms more readily than it catches. We measure all three.
So an AI reviewer that nobody has measured does not remove the validation burden. It hides it.
The shape that actually works
Seventeen of our nineteen instrumented failure modes are scored by rule — resolve the DOI, validate the schema, replay the ledger, check the units.
Why: Rules scale like generation does. They cost nothing per invocation and they do not drift when a vendor ships an update.
Let a model rank what a human should look at first. Do not let it decide what a human can skip.
Why: Our own triage tool re-checks each failure against the same code path the detector used and approves nothing. It was one commit away from auto-approving 'obvious' cases, which would have been the bypass the gate exists to prevent, wearing a helpful name.
We bulk-accepted 2,902 failures on one run. Every probe carries reviewed_by: bulk-accept in the data.
Why: A validation burden you have decided not to carry is fine. One you have quietly stopped carrying is not, and the difference is entirely whether the record says so.
The timeline
Immediate: review queues grow faster than teams and quality gates get satisfied nominally.
Near-term: AI reviewers are deployed to absorb it, mostly unmeasured, and the first class of incident appears where the reviewer and the generator failed the same way on the same input.
Dominant: the organisations that scaled deterministic checking rather than model-based checking are the ones whose throughput is actually usable.
What would prove this wrong: a measured demonstration that model reviewers do not correlate with the generators they check. I have not seen one. If you have, I genuinely want it.
What proportion of your AI output actually gets validated — and is that number going up or down?
Every figure here describes something measured and committed. See the measurements · read the method