When AI Makes Research Cheaper, Verification Can Become the Bottleneck

A problem beyond submission volume

Generative AI can reduce the effort involved in writing, coding and developing research. But does easier production necessarily mean more established knowledge? Peter Achim’s November 2026 preprint, The Verification Bottleneck, develops an economic model in which the answer is no.

The argument is not simply that more manuscripts overwhelm journals. The model keeps the researcher population fixed. Instead, cheaper production changes how effectively editors can distinguish research of different underlying quality, and therefore how productively they can allocate scarce expert attention.

This is a theoretical result, not an empirical demonstration that AI has already reduced verified scientific knowledge.

How research scope becomes a signal

The model has three groups: researchers, an editor and reviewers. Researchers privately know their project’s quality or promise. They choose an observable research scope—the breadth and ambition of the questions and claims they develop. The editor commits to assigning review depth according to that scope.

Two distinctions matter. Review depth is the intensity of scrutiny; verification load is the reviewer time needed to carry it out. A broader project can require more time even if examined less deeply. “Verified value” means the expected scientific contribution established through review, rather than a count of papers or claims.

Scope also helps researchers reveal their quality through their choices. Stronger researchers are more willing to take on additional scope relative to the scrutiny they receive. Demanding assignments can therefore help separate them from weaker researchers.

But screening has a cost. Under the conditions Achim studies, the strongest research receives less scrutiny than its marginal scientific return would justify, while the weakest receives more. Stronger researchers also face scope requirements beyond the format that would minimize reviewer time for their assigned verified value. Review is doing two jobs: establishing knowledge and keeping researchers willing to choose their intended assignments.

Why cheaper production undermines screening

When technology lowers the cost of developing scope, ambitious assignments become easier to imitate. A scope requirement that once discouraged a weaker researcher loses some of its deterrent effect.

The editor can respond by demanding still broader work from stronger researchers. Yet that work must itself be checked. Achim shows that preserving a fixed benefit from imitation, while continuing to establish positive scientific value from the stronger assignment, eventually requires more reviewer time than any finite budget allows.

An optimizing editor therefore has to reorganize assignments, narrow their benefits or pool researchers into common assignments. Beyond a technological threshold, the model predicts that a fixed stock of reviewer time produces no more verified knowledge as production becomes cheaper—and can produce less. The threshold is a mathematical feature of the model, not an estimated date or AI capability level.

Reviewer withdrawal adds another channel

Reviewer capacity is not fixed in the full model. Reviewers volunteer time according to its expected scientific contribution, weighed against the opportunity cost of reviewing. They decide before knowing which particular submission they will receive.

If the expected contribution per reviewer hour falls, they supply less time. That withdrawal can amplify the initial loss in verification productivity.

Under the paper’s assumptions, both reviewer capacity and verified knowledge eventually become non-increasing as production technology improves. They need not fall with every improvement: pooling can produce a plateau. Additional conditions establish strictly declining outcomes when differentiated assignments remain productive but screening remains costly.

Multiple equilibria are possible. The headline comparison follows the equilibrium with the greatest reviewer capacity, which also has the most verified knowledge. Another central assumption holds the marginal opportunity cost of reviewing unchanged as production technology advances. Researchers and reviewers are modeled as separate populations.

If editors could observe quality directly, the result reverses: cheaper production would leave reviewer capacity and verified knowledge unchanged or higher.

Remedies—and the welfare distinction

Achim identifies two ways to improve outcomes within the framework. Better verification technology reduces the time needed to establish scientific value. More informative preliminary screening reduces reliance on costly scope requirements. Both can increase attainable verified knowledge and support greater reviewer capacity.

Scope-dependent submission fees offer another screening instrument. In the paper’s two-type illustration, suitable fees can restore the full-information allocation when production costs are sufficiently low. This is not a general result for ordinary flat submission fees.

A separate market benchmark lets researchers pay for the reviewer resources they use. Cheaper production then raises maximized surplus net of production and verification costs, although the paper stresses the difficulty of contracting for review quality.

Finally, less verified knowledge does not automatically mean lower overall welfare. Production-cost savings also matter, leaving the effect on total surplus ambiguous. The preprint’s central warning is narrower: improving research production alone can weaken scientific verification when institutions depend on increasingly cheap signals to allocate scrutiny.