Peer review rests on a single scarce resource: the unpaid attention of working scientists, and that resource is being rationed thinner every year. Publications indexed in Scopus and Web of Science are growing at roughly 5.6 percent per year, a rate some analyses describe as exponential. Estimates put the collective time researchers spend reviewing at around 15,000 person-years annually, work that would cost about $1.5 billion if paid at market rates in the United States alone. The supply of expert attention is not growing with the demand, and the AI era is making the imbalance worse.
The cost of the squeeze shows up in individual cases. Jason Semprini, a health economist at Des Moines University, submitted a paper on state policies that require elementary school students to receive the HPV vaccine. His finding was counterintuitive but specific: the mandates did not dramatically reduce cervical cancer rates at the population level, largely because some families respond to mandates by avoiding the vaccine altogether. A reviewer read the submission and concluded Semprini was questioning whether the HPV vaccine prevents cervical cancer at all. On that misunderstanding, with a single reviewer assigned, the paper was rejected.
The numbers behind the squeeze
The arithmetic of peer review has been deteriorating for years, but the pieces are now visible in the data. Journal submission volumes have climbed across disciplines. Reviewers, who typically work anonymously and without pay, report being asked to review more papers than they can reasonably handle, and declining invitations at higher rates. The result is a system in which journals sometimes send a manuscript to a single reviewer, as happened to Semprini, rather than the two or three that would once have been standard. One misunderstanding, one rushed reading, one overworked expert, and the gate closes.
The unpaid labor is the system’s structural weakness. Researchers review because it is part of the academic contract, because they benefit from the same service, because they want to shape their field. But the incentive is diffuse and the cost is concrete. Every review is time taken from the reviewer’s own research, and when submission volumes rise while the reviewer pool stays flat, quality is what bends.
The AI amplifier
Artificial intelligence is entering this fragile system from two directions at once. On the production side, AI-assisted drafting makes it faster and cheaper to write and format a manuscript, lowering the barrier to submission and raising the number of papers journals must process. The underlying science may be perfectly sound, but the volume problem is worsened by the ease of production.
On the evaluation side, AI is being used to generate reviews themselves, often poorly. At the ICLR 2026 machine-learning conference, an analysis of more than 75,000 reviews found roughly 21 percent appeared to be fully AI-generated, with over half showing at least some signs of AI assistance. At ICML 2026, 506 reviewers were caught violating AI-use policies, detected through embedded watermarks, and nearly 500 papers were desk-rejected, about 2 percent of submissions. The problem is not confined to computer science. Paper mills, commercial operations that mass-produce fake research for paying authors, have industrialized fabrication, and an analysis of cancer literature published in Nature estimated that roughly one in ten papers in the field may originate from such mills. Fabricated citations have been found in published work, including a Lancet research letter reporting the problem in a fraction of a percent of indexed papers.
The danger is not that every AI-assisted paper is fraudulent. It is that reviewer bandwidth, not scientific merit, becomes the limiting resource determining which claims receive serious scrutiny. When one overworked volunteer stands between a manuscript and the literature, the probability that a genuine misunderstanding sinks good work rises, and the probability that a fabricated or sloppy paper slips through rises with it.
What is being tried
The system is not without responses. Journals are experimenting with paying reviewers, with mixed results reported by publishers who have run trials. Nature has been extending transparent peer review, publishing reviewer reports alongside papers, a change associated with lower retraction rates. Some venues are testing applicant cross-review, where authors review one another’s submissions, and a growing number of funders now use randomization to break ties among proposals judged equally meritorious, reviving the ancient Athenian kleroterion as a tool of fair allocation.
AI tools are also being developed to assist, not replace, reviewers: screening submissions for fabrication indicators, checking references, flagging suspicious statistics. The position that AI should not yet write reviews outright is gaining ground in the machine-learning community itself, where empirical comparisons of human and AI-generated reviews have shown the gap in quality and reliability.
The structural choice
None of these fixes addresses the underlying equation. Science produces more output every year, journals compete to publish quickly, researchers need publications for careers, and AI is making written production easier. The one input that cannot be scaled at will is expert attention. Peer review depends on a pool of working scientists donating time they increasingly do not have, and no amount of process redesign changes that bottleneck.
The likely next phase of the debate will be about whether the answer is compensation, triage, AI assistance with human oversight, or some redesign of what review is for. What the Semprini case shows is that the question is not abstract. A paper that could have contributed to a genuinely important policy debate, about whether vaccine mandates work as intended, was held back by a single misreading, and the researcher’s experience is increasingly common across disciplines.
Policymakers, doctors, investors and the public treat journal publication as a proxy for legitimacy. If the process behind that proxy is thinning, the signal becomes noisier at exactly the moment when trust in expertise is already contested. The system that filters bad science from the record is itself in need of a filter, and the AI era is forcing the question faster than the institutions can answer it.
Sources:
- Ars Technica, “Peer review is overwhelmed, can it survive in the AI era?” (Aug 10, 2026): https://arstechnica.com/science/2026/08/peer-review-is-overwhelmed-can-it-survive-in-the-ai-era/
- Nature News Feature, “The peer-review crisis: how to fix an overloaded system” (Aug 2025): https://www.nature.com/articles/d41586-025-02457-2
- Journal Metrics analysis of ICLR 2026 AI-generated reviews (Pangram Labs data): https://www.journalmetrics.org/blog/ai-written-peer-reviews-medical-authors-2026
- PeerJ / Research Integrity & Peer Review (reviewer workload estimates): https://link.springer.com/article/10.1186/s41073-021-00118-2
- arXiv, “Stop Automating Peer Review Without Rigorous Evaluation”: https://arxiv.org/abs/2605.03202