AI Peer Review Works — But Only If You Understand the Market

Think about what peer review actually is, stripped of its academic prestige: it's a quality-certification system that relies almost entirely on volunteer labor, has no real price mechanism to balance supply and demand, and is structurally incapable of scaling with the volume of product it's supposed to certify. Journals charge authors (or their institutions) submission fees and charge readers (or their libraries) subscription fees, while the certifiers — peer reviewers — receive nothing except the vague social capital of being seen as a collegial participant in the knowledge enterprise.

That model was already under pressure before AI entered the picture. Now AI has turbocharged manuscript production while doing almost nothing for reviewer supply. As economist Tyler Ransom noted on his Substack, writers including Scott Cunningham, Paul Goldsmith-Pinkham, and Brian Heseung Kim have all flagged how AI is accelerating the research process and what that acceleration means for a peer review infrastructure that wasn't built to absorb it. The system's gatekeepers are being overwhelmed by the gate.

So when researchers start testing whether AI can step in as a reviewer, the relevant question isn't just "does it work?" It's: who designed this system, what were their incentives, and what happens to those incentives when you plug AI into the certification function?

What the experiment actually found

A recent experiment described at Marginal Revolution provides the sharpest empirical data point available on this question. The researcher — identified via a related Marginal Revolution post as Brian Chau — deliberately planted 100 known errors into 10 open-access psychology papers, then ran them through frontier AI models and two commercial AI review tools to see what each system would catch.

The headline number is the spread: the best single system caught 71 of 100 errors. The worst caught 30. That is not a rounding difference — that is a gap wide enough to matter enormously depending on which tool a journal happens to deploy, which field it covers, and what type of errors are most common in its submissions.

If you're accustomed to thinking in terms of risk distributions rather than point estimates, that 30-to-71 range should immediately concern you. A single model with a 30% catch rate isn't a peer reviewer — it's a false-negative machine that provides the appearance of oversight without the substance. It's the academic equivalent of a credit rating that conveys AAA confidence while the underlying assets are deteriorating. The certification function gets performed; the quality function doesn't.

What changes the picture is pooling. When Chau combined results across multiple systems, 93 of 100 errors were detected. That number is genuinely striking — and it's the figure that should anchor any serious conversation about AI's role in this space. It suggests that the value of AI in peer review isn't in replacing individual expert judgment with a single model. It's in running parallel systems and flagging where they disagree or where any one system raises a concern, which is a structurally different product than what any single tool currently offers.

The incentive problem the accuracy numbers can't solve

Here's where the market framing matters more than the benchmark scores.

The Committee on Publication Ethics (COPE), whose guidance on AI tools is available at their site, concluded that "it is unlikely that any current AI can reliably carry out a well-substantiated review" — and recommended thinking about AI as a tool to improve human-authored reports and to perform triage tasks, not as a standalone certifier. That framing — triage tool alongside human reviewers — makes sense on quality grounds. But triage is cheaper and faster than full review, which means the economic pressure on journals will be to use AI to do more of the work with fewer humans, not to use AI to make human review more rigorous.

That pressure isn't hypothetical. Publishers operate in a commercial environment. Article processing charges at major open-access journals now run into thousands of dollars per paper. The throughput model — volume of published papers — is baked into how journals, and by extension universities, measure productivity. The incentive to process submissions faster, at lower per-unit cost, runs directly against the incentive to invest more deeply in quality verification.

As an arXiv preprint on AI and the Future of Academic Peer Review observed, peer review has traditionally served functions beyond error-catching: creating moments of dialogue, mentorship, and recognition between early-career and established researchers. If AI takes on a larger evaluation role, those functions erode. That loss is real, but it's also the kind of loss that doesn't show up in any publisher's cost-benefit analysis — which is precisely why it tends to get discounted in institutional decisions.

The researchers in an arXiv study examining what happens when reviewers receive AI feedback observed a similar dynamic: advocates see AI's potential to reduce reviewer burden and improve quality, while critics warn of risks to fairness, accountability, and trust. Both can be right simultaneously, and usually are, when the technology is real but the governance isn't.

Who actually has skin in this

It's worth being specific about where the economic incentives converge and diverge.

Commercial publishers benefit from AI that reduces reviewer recruitment costs — finding willing, qualified reviewers is now one of the most labor-intensive parts of journal operations. They have less direct incentive to use AI in ways that increase rejection rates or slow throughput, because both hurt revenue. The interests of publishers and the interests of scientific quality are not automatically aligned; they sometimes are, but assuming alignment is how you end up with a broken certification system.

Funders — research councils, foundations, government science agencies — have stronger incentives to care about quality, because their money is effectively downstream of peer review. A paper that gets published through a degraded review process and later retracted represents wasted grant money. Several major funders have begun requiring data sharing and pre-registration precisely because they don't fully trust the existing quality infrastructure. AI that is genuinely rigorous serves their interests; AI that provides cover for faster publication does not.

Universities face a structural contradiction. Their researchers need publications for promotion and grant applications, which creates pressure to produce volume. But their reputations depend on quality, and institutions whose researchers publish retracted or unreliable papers pay reputational costs. Whether AI in peer review helps or hurts universities depends on exactly how it gets implemented — and by whom.

Researchers themselves, particularly early-career ones from institutions with less access to peer networks, may actually benefit most from well-calibrated AI review. As The Conversation's analysis of AI's risks to peer review quality notes, current guidance on AI use in peer review is mixed and its effectiveness unclear — but the researchers most likely to receive "vague, confusing" AI feedback that does nothing to improve their work are disproportionately those who don't have senior mentors to compensate for the gap.

What the spread tells you about what to build

Return to that 30-to-93 range. The 30% figure is a floor produced by a single, apparently weak model working alone. The 93% figure is a ceiling produced by aggregating multiple models. That spread isn't just a performance gap — it's a design specification for what an actually useful AI peer review system would look like: diverse model inputs, aggregated signals, human review of what the models flag as uncertain or contradictory.

The problem is that designing a system like that requires up-front investment in architecture, not just a subscription to a commercial tool. Journals that treat AI review as a cost-cutting measure — one tool, deployed at scale, with minimal human oversight — will cluster toward the 30% end of that distribution. The certification function will be performed; the quality function won't be. The paper will say "AI-assisted review." The error will still be in print.

The gap between those two outcomes is not a technology question. It's a question of who is paying for what, and why.

---

Raj Mehta covers international finance, sovereign debt, and the global economy for BuzzRAG.